Leveraging domain knowledge and artificial intelligence to facilitate data driven decision making

A domain knowledge-based AI system addresses the challenge of deterministic and traceable responses to legal and regulatory queries, providing reliable insights by tracing responses to verified sources, enhancing decision-making in financial and regulatory contexts.

WO2026236712A1PCT designated stage Publication Date: 2026-11-12DATARIGHTS AI INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/027453
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-09
Filing Date
2026-05-08
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing AI systems struggle to provide deterministic and reproducible responses to legal and regulatory queries, often relying on statistical inference and lacking traceability to verified source material, which can lead to inaccurate and potentially harmful outputs.

Method used

A computer-implemented method using a domain knowledge base curated by experts to process queries, providing plain-language responses with traceable citations, leveraging a hybrid AI system that includes a small language model to ensure deterministic and rule-based processing across structured domain documents.

Benefits of technology

Enables reliable and traceable responses to complex legal and regulatory queries, ensuring accuracy and reliability by tracing each step to verified source documents, reducing the risk of inaccuracies and enhancing decision-making in financial and regulatory contexts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026027453_12112026_PF_FP_ABST
    Figure US2026027453_12112026_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed are systems, methods, and computer program products for document intelligence. The disclosed systems, methods, and computer program products include a curated domain knowledge base comprising a knowledge graph constructed during expert-led training sessions, a query-processing architecture that returns plain-language responses. Also disclosed are training methods for generating the domain knowledge base from structured domain documents.
Need to check novelty before this filing date? Find Prior Art

Description

LEVERAGING DOMAIN KNOWLEDGE AND ARTIFICIAL INTELLIGENCE FOR EXTRACTING DATA AND INTELLIGENCE FROM LEGAL, REGULATORY AND FINANCIAL TRANSACTIONS AND WORKFLOWS TO FACILITATE DATA DRIVEN DECISION MAKINGCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 803,389, filed May 9, 2025. The entire contents of the above-identified applications are hereby fully incorporated herein by reference.TECHNICAL FIELD

[0002] The subject matter disclosed herein is generally directed to artificial intelligence systems and methods for structured document intelligence, including systems and methods that apply expert-curated domain knowledge and industry context to process one or more user queries against a structured domain document library and return plain-language responses with traceable citations to source documents. In addition, the systems and methods disclosed herein include training methods for generating domain knowledge bases through expert-led curation sessions, query-response architectures applicable across structured document domains, and multi-document processing architectures configured to simultaneously resolve a vast plurality of queries across a vast plurality of structured domain documents.BACKGROUND

[0003] Information is a critical business asset with significant potential value. While entities like financial institutions seek to understand and aggregate the information contained in comprehensive and voluminous libraries of proprietary written documents, and regulations, they struggle with navigating competing and conflicting regulations and regulatory documentary requirements and interpretations on a global scale. Institutions also struggle with understanding and aggregating the contents and interpretations of legal contracts and related documents, so that the information can be mined to yield actionable insights for compliance, corporate governance and product development and operational efficiency. For example, licensees of proprietary intellectual property may misunderstand or not know their usage rights to such proprietary intellectual property, these licensees may need to interrogate source documents such as regulations and their own proprietary and confidential documents and agreements in order to obtain the information that they seek to understand, and they may further need to comprehend that information in the context of a highly regulated business environment.In addition, their business record libraries of their proprietary and confidential legal contracts and documents and their regulations change frequently and in each case the business is required to adapt to new information dynamically such as when new contracts are executed and new regulations are approved.

[0004] Financial institutions and regulated-industry entities frequently maintain document libraries comprising thousands of contracts, regulatory instruments, and related agreements. Obtaining actionable intelligence from such a library requires the ability to process large numbers of documents simultaneously across multiple concurrent queries, with each response traceable to its underlying source. A need exists for systems capable of operating at corpus scale in this manner while preserving per-response, passage-level traceability across the full document library.

[0005] Legal and regulatory analysis requires that system-generated responses be deterministic and reproducible, derived from defined reasoning steps tied to identified source content rather than from statistical inference. There remains a need for Al -based systems whose response logic can be traced through a defined sequence of steps to verified source material, and that do not rely on probability -based computation to generate responses.

[0006] Modem Al systems could enhance the productivity of legal, regulatory and operational professionals if a LLMs' statistical probability-based and pattern matching computations can be improved. More specifically, a need exists for an approach in which response logic is deterministic and derived from expert-curated domain knowledge, rather than from statistical computation, such that every system-generated response can be traced through a defined sequence of steps to verified source material. At present, LLMs can produce problematic hallucinations and occasional guesswork, potentially with disastrous outcomes.

[0007] There further remains a need for Al systems operating in legal, regulatory, and compliance contexts to provide full response traceability: verification that each systemgenerated response is supported by a specific, identified passage in an identified source document, and a defined negative output state when no document in the corpus is responsive to the query posed.

[0008] Citation or identification of any document in this application is not an admission that such a document is available as prior art to the present invention.SUMMARY

[0009] In an aspect, the techniques described herein relate to a computer-implemented method for analyzing documents, comprising: receiving, from a user, one or more queries associated with a document library comprising a plurality of structured domain documents; processing, by an Al system or hybrid Al system, the one or more queries using a domain knowledge base curated by one or more domain experts; and providing, to the user, a response to the one or more queries, wherein the response is in plain language and comprises a reference identifying one or more of the documents relied upon to generate the response. In an embodiment, the Al system or hybrid Al system comprises one or more machine learning models comprising one or more of a large language model (LLM), a small language model (SLM), a topic model, a Latent Dirichlet Allocation (LDA) model, a neural network, or a combination thereof. In an embodiment, the Al system or hybrid Al system comprises a small language model (SLM). In an embodiment, the reference comprises an interactive reference to one or more of the documents, one or more passages within the one or more documents relied upon to generate the response, or a combination thereof. In an embodiment, processing the one or more queries comprises transforming the one or more queries into one or more enriched queries by aligning terminology of the one or more queries with domain-specific language of the domain knowledge base. In an embodiment, the one or more queries comprise free-form queries, queries selected from a domain-specific query library associated with the domain knowledge base, or a combination thereof. In an embodiment, the domain knowledge base comprises a knowledge graph or other symbolic representation of knowledge comprising a plurality of nodes and a plurality of edges. In an embodiment, the knowledge graph or other symbolic representation of knowledge is a network graph. In an embodiment, the network graph comprises a hierarchical graph, a directed acyclic graph, a property graph, an ontology graph, a semantic network graph, a multi-relational graph, a hybrid thereof, or a combination thereof. In an embodiment, the plurality of nodes represent one or more document types, agreement provisions, statutes, statutory provisions, regulatory frameworks, compliance requirements, risk factors, jurisdictions, entities, concepts, or a combination thereof. In an embodiment, the plurality of edges represent relationships between the nodes. In an embodiment, the plurality of edges independently comprise one or more of a supersession relationship, a section relationship, a read-for relationship, a related-to relationship, a jurisdictional relationship, a compliance relationship, a risk relationship, a cross-reference relationship, a level relationship, or a combination thereof. In an embodiment, the processing comprises traversing the knowledge graph or other symbolic representation of knowledge toretrieve one or more portions of the domain knowledge base relevant to the one or more queries. In an embodiment, traversing the knowledge graph or other symbolic representation of knowledge comprises traversing from an organizational level to a section level to a clause level to a context level. In an embodiment, the method further comprises receiving, from the user, a modification to one or more nodes or edges of the knowledge graph or other symbolic representation of knowledge, and updating the knowledge graph or other symbolic representation of knowledge to reflect the modification. In an embodiment, the processing is deterministic. In an embodiment, the method further comprises generating a reasoning chain tracing each step taken to generate the response. In an embodiment, the method further comprises providing, to the user, an indication of one or more processing steps being performed during the processing. In an embodiment, processing the one or more queries comprises processing the one or more queries simultaneously against a plurality of documents in the document library. In an embodiment, the plurality of documents comprises at least 10 documents, at least 100 documents, at least 1,000 documents, at least 10,000 documents, or at least 100,000 documents. In an embodiment, the response comprises a matrix, a summary of the response, a visualization of the response, or a combination thereof. In an embodiment, the matrix comprises a plurality of cells, each cell corresponding to one or more documents of the plurality of documents. In an embodiment, each cell of the matrix comprises a plain-language response for the one or more documents corresponding to that cell. In an embodiment, each cell of the matrix comprises an identification of one or more passages within the one or more documents corresponding to that cell relied upon to generate the response for that cell. In an embodiment, rows and columns of the matrix are transposable. In an embodiment, one or more rows of the matrix each correspond to one or more of the one or more queries, and one or more columns of the matrix each correspond to an organizational level, a section level, a clause level, a context level, or a combination thereof. In an embodiment, each cell of the matrix is individually inspectable to display one or more source documents and a reasoning chain underlying the response for that cell. In an embodiment, the response further comprises score types associated with each of the one or more passages relied upon to generate the response. In an embodiment, the visualization of the response comprises a domain-specific visualization corresponding to a type of domain associated with the one or more queries. In an embodiment, the visualization of the response comprises one or more nodes and edges of the knowledge graph or other symbolic representation of knowledge traversed during the processing. In an embodiment, the method further comprises, when the domain knowledge base does not contain information responsive to the one or more queries, returning an unsure response rather thangenerating an answer. In an embodiment, the method further comprises displaying a document viewer. In an embodiment, the document viewer displays one or more documents, one or more passages within the one or more documents, or a combination thereof relied upon to generate the response. In an embodiment, the one or more passages are highlighted within the document viewer. In an embodiment, the domain knowledge comprises domain-specific terminology for one or more document types, cross-reference mappings between related agreement sections, document priority hierarchies, or a combination thereof. In an embodiment, the document library is stored in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users. In an embodiment, the processing comprises: aligning query terminology with domain-specific language using an ontology and definitions module; refining the one or more queries against the domain knowledge base using a question enrichment module; selecting relevant documents from the document library using a document filtering module; traversing the domain knowledge base from an organizational level to a section level to a clause level to a context level to identify relevant portions of the domain knowledge base using a traversal module; segmenting the relevant portions into text segments using a chunk extraction module; resolving clause hierarchies and cross-references using an ordering and document relationships module; compiling and composing a coherent response using a synthesis module; validating logical soundness of the compiled response against the domain knowledge base using a reasoning module; evaluating relevance of domain knowledge relied upon to generate the response using a scoring module; or a combination thereof. In an embodiment, one or more of the ontology and definitions module, the question enrichment module, the document filtering module, the traversal module, the chunk extraction module, the ordering and document relationships module, the synthesis module, the reasoning module, or the scoring module operate according to a document-type-specific configuration template that governs the order and manner in which the one or more modules are invoked for a given document type. In an embodiment, the method further comprises receiving, from the user, one or more additional queries and repeating the processing and providing steps for the one or more additional queries. In an embodiment, the structured domain documents comprise contracts, legal documents, statutes, regulatory documents, financial documents, compliance documents, or a combination thereof. In an embodiment, the contracts, legal documents, the financial documents, or a combination thereof comprise commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, government contracts, securities filings, loan agreements, or acombination thereof. In an embodiment, the regulatory documents, the compliance documents, or a combination thereof comprise regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions, or a combination thereof. In an embodiment, the method further comprises storing the response on a local storage device, a cloud-based storage service, or a combination thereof. In an embodiment, the method further comprises transmitting the response to another user, a computing system, or a combination thereof. In an embodiment, the domain knowledge base is generated during a plurality of training sessions conducted by one or more domain experts using data associated with the user, domain-specific curated data, or a combination thereof.

[0010] In an aspect, the techniques described herein relate to a method of generating a domain knowledge base for use in analyzing queries associated with documents, comprising: assembling reference data comprising one or more documents and document-type-specific information associated with the one or more documents; and curating domain knowledge from the reference data during one or more training sessions conducted by one or more domain experts to generate a domain knowledge base. In an embodiment, curating the domain knowledge comprises generating a knowledge graph or other symbolic representation of knowledge comprising a plurality of nodes representing one or more document types, agreement provisions, statutory provisions, regulatory frameworks, compliance requirements, risk factors, jurisdictions, entities, concepts, or a combination thereof, and a plurality of edges representing relationships between the nodes. In an embodiment, the plurality of edges independently comprise one or more of a supersession relationship, a section relationship, a read-for relationship, a related-to relationship, a jurisdictional relationship, a compliance relationship, a risk relationship, a cross-reference relationship, a level relationship, or a combination thereof. In an embodiment, curating the domain knowledge comprises encoding document priority hierarchies that define an order of precedence among two or more of the one or more documents. In an embodiment, curating the domain knowledge comprises disambiguating domain-specific terminology by associating one or more terms with contextspecific meanings based on the document type in which each term appears. In an embodiment, curating the domain knowledge comprises mapping cross-references between related agreement sections across two or more of the one or more documents. In an embodiment, the method further comprises conducting one or more additional training sessions to extend or update the domain knowledge base following an initial set of training sessions. In an embodiment, the method further comprises generating a document-type-specific configuration template for one or more document types, the document-type-specific configuration templategoverning the order and manner in which one or more processing modules are invoked when analyzing a document of that document type. In an embodiment, the one or more domain experts have expertise in commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, statutes, regulatory filings, compliance policies, merger and acquisition due diligence, taxation, government contracts, securities filings, loan agreements, statutory provisions, internal controls, government regulatory submissions, or a combination thereof. In an embodiment, the reference data comprises contracts, legal documents, regulatory documents, financial documents, compliance documents, or a combination thereof, comprising one or more commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, government contracts, securities filings, loan agreements, regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions, or a combination thereof. In an embodiment, the domain knowledge base is structured to support deterministic, rule-based processing of queries. In an embodiment, the method further comprises storing the domain knowledge base in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users. In an embodiment, the method further comprises establishing one or more scoring parameters or relevance thresholds for evaluating domain knowledge relied upon during processing of queries against the domain knowledge base. In an embodiment, the reference data comprises data associated with the user, domain-specific curated data, or a combination thereof.

[0011] In an aspect, the techniques described herein relate to a system to analyze documents, comprising: a storage device; and a processor communicatively coupled to the storage device, wherein the processor executes application code instructions that are stored in the storage device to cause the system to: receive, by one or more computing devices, one or more queries from a user, the one or more queries associated with a document library comprising a plurality of structured domain documents; process, by an Al system or hybrid Al system, the one or more queries using a domain knowledge base curated by one or more domain experts; and transmit, by the one or more computing devices, a response to the one or more queries to a device associated with the user, wherein the response is in plain language and comprises a reference identifying one or more of the documents relied upon to generate the response.

[0012] In an aspect, the techniques described herein relate to a computer program product, comprising: a non-transitory computer-readable storage device having computer-executable program instructions embodied thereon that when executed by a computer cause the computer to analyze documents, the computer-executable program instructions comprising: computerexecutable program instructions to receive, by one or more computing devices, one or more queries from a user, the one or more queries associated with a document library comprising a plurality of structured domain documents; computer-executable program instructions to process, by an Al system or hybrid Al system, the one or more queries using a domain knowledge base curated by one or more domain experts; and computer-executable program instructions to transmit, by the one or more computing devices, a response to the one or more queries to a device associated with the user, wherein the response is in plain language and comprises a reference identifying one or more of the documents relied upon to generate the response.

[0013] Within this framework, a need has risen that overcomes these and similar deficiencies from effectively leveraging today’s technology for determining how financial firms and other institutions extract critical and reliable contract information, delivering instant and actionable insights for facilitating informed decision-making.

[0014] More specifically, systems, methods, and platforms are introduced herein that embed expert-domain knowledge in a manner, offering an unparalled value proposition that currently does not exist in today’s market. This differentiation has been valued through extensive interviews with Proof of Concept (POC) clients, potential customers, and significant market research.

[0015] The techniques of the present disclosure recognized the problem that companies and other business entities must first have a clear understanding of the rights and obligations embedded in their contracts. While a variety of applications use cases naturally emerge from this technology, and fall within the scope of the principles introduced herein, an exciting area of focus is the extraction of important information in a consistently reliable way. For one, data is the subject of insatiable demand on Wall Street and other areas.

[0016] Accordingly, in one aspect of the present disclosure, the rights and obligations in contracts can be unlocked instantaneously, and in a highly reliable manner.

[0017] The new technology introduces techniques for using domain expert knowledge and a specialized Al system to answer high-stakes questions for users related to complex legal documents in financial, or a host of other services. The domain of financial services may include, for example, market data and related transactions and contractual relationships. In anembodiment, alternative legal issues may be used in the context of such systems, including complex legal matters like due diligence in M&A transactions, regulatory filings, compliance policies, taxation and others.

[0018] In one aspect of the disclosure, an artificial intelligence (Al) enhanced system includes at least one processor configured to receive queries and to provide responses to the queries based on a user’s proprietary and confidential reference data, memory coupled to the one or more processors, the memory including code for execution by the one or more processors, the one or more processors further coupled to another memory configured to store the reference data during a plurality of training sessions conducted by a legal professional using at least one input device coupled to the first computing device, a large language model (LLM) coupled to the first computing device to perform ancillary tasks during operation of the first computing device without affecting integrity of the training data, wherein the training data comprises knowledge used by the one or more processors to respond to the queries via a user interface related to legal documents in financial services.

[0019] The specialized Al system may provide access to its domain knowledge using a query-response format to receive a prompt and provide the user with relevant responses based on pertinent documents, significant agreement provisions, key concepts and relationships between these objects. In an embodiment, the relevant transactions may include market data contracts of a financial institution, due diligence in a transaction, regulatory filing or other types of document analysis.

[0020] To facilitate the capabilities of the disclosed system according to one aspect of the disclosure, reference data may initially (and subsequently periodically) be uploaded to a computer to generate, augment, update or otherwise revise a very high reliability model. Unlike conventional approaches, the data is not merely gathered from a variety of potentially-relevant (and possibly inaccurate) sources. Rather, data in the relevant domain is collected based on the practical and real-world knowledge and experience of one or more persons deemed experts in the pertinent fields of business data exploitation, market data contracts and related legal transactions, regulatory compliance issues, and the legal and practical effects of each. In other cases, the consulted legal experts are much more specialized. The experts, including legal professionals such as highly experienced attorneys in these fields, may have significant, longstanding experience in facets of data science that include (among other detailed knowledge) the features, common documentary provisions, important legal definitions, and relevant legal instruments needed to exploit datasets of a business entity. The training knowledge from the experts may include the specific legal and practical considerations anddetails relevant to a particular use case together with specific means to organize and execute the often copious documentary instruments required to consummate a transaction.

[0021] In another aspect of the disclosure, the received expert knowledge may be curated to enable the specialized Al system to provide pertinent responses to user queries. In addition to the specially-tailored responsive information, the responses may give the user specific access to the documentary sources provided by the legal professional that serve as the factual or legal basis for the responses. In an embodiment, the system highlights the specific provisions relied on within the documentary sources, so that a subsequent user can focus attention on those provisions without having to read the entire excerpt.

[0022] Still further to the documentary sources, the curated knowledge may include a plurality of key concepts subsequently referenced in a user query. The identified key concepts may be based on the user query and relevant to the specific actions to be undertaken. Thus, in one embodiment, the knowledge is highly selective to integrate the necessary tools, including the documentary identities or agreement excerpts, with the plurality of key concepts that relate to the document types and the user’s stated objective in the response to the query. The domain knowledge is also used to explain to the user in a concise and readable format the nature of the information set forth in the response to the query. The specific key concepts present from the training phase may be initially formulated by a user query, ensuring a relevant response and a display of the reasoning used by the code ranked.

[0023] One overarching objective of the domain knowledge and key concepts is to enable the user of the Al system to have a reliable source for specifically-delineated terms and conditions in an instant.

[0024] In addition to the document references and key concepts, the reference data may also include information concerning the relationships between different document types and / or key concepts. The relationship data may be used initially by the professional during training sessions to link different types of textual or other such objects together, while providing information about the relationship. Thus, this relationship information may be initially obtained from the expert, whether directly, or based on relevant sources relied on by the expert in the training phase. In this vein, the system can provide a body of domain knowledge along with specified relationships between different documents, between documents and key concepts, or between two or more key concepts for providing comprehensive responses to user inquiries. The relationship information may assist the user in understanding the overall flow of the desired transaction. This information may also help assure that the levels of precision and accuracy needed are integrated by the user into the related legal instruments.

[0025] While the curation may be performed using a variety of techniques, different embodiments enable the use of visual mechanisms that rely on clusters of nodes or objects formed based on the user query. The clusters may be linked to and displayed over UIs and touchscreens. These tools can accord the user a distinct advantage in visually understanding the data during a query session. The tools may include multiple visual and textual-based tiers that the user can explore for more information by selecting an object or icon. The user can receive comprehensive responses supported by logical sources and clarifications supplied by the system. With the system using a verbal or visual-based tool, described further herein, the specialized Al system is invaluable for educating the user and informing the user of most or all facets of a desired transaction or other event.

[0026] In particular, the domain knowledge may be represented by a knowledge graph or other algorithmic representation of human comprehension ( knowledge graph) as referred to herein by the inventors. The knowledge graph represents a hierarchical set of manipulable information symbols or objects that can visually represent, responsive to a query, a cluster of nodes, with links defining relationships between specific nodes as noted above. The hierarchical nature of the knowledge graph allows a user to make selections of different nodes or accompanying text to review an underlying segment of information regarding a selection by a user of a node in the cluster. Depending on the implementation, further selections may be made that may include the relevant documentary sources supporting the query, an identification of the key passages within the documentary sources, clarifications of the text at issue from the trainer, and further access to related information.

[0027] The reference data is elicited from the expert and from documentary sources generated or relied on by the expert, part or all of which can be stored in a non-volatile memory such as a solid state drive (SSD), a magneto-based drive, NAND orNOR-based flash memory, programmable read only memory (PROMs) or other nonvolatile storage hardware. In an embodiment, the training data can be modified, supplemented, deleted, or otherwise augmented during both the initial training and the operational phases of the system. Likewise, a user can customize the data “on the fly” and can create, in or near real-time, modified models specific to the needs of the business entity accessing the system.

[0028] In an embodiment, the knowledge graphs may be manipulated, such as by zooming in to identify a particular set of nodes of interest. In an embodiment, the substance of the knowledge graphs (e.g., the node cluster) may be modified and tailored to the needs or practices of a particular financial institution or user. Further, as noted in an embodiment, the node cluster may be part of a hierarchical multi-layered data structure. For example, the user may selectdifferent portions of the knowledge graph to identify not only the sources from which the documents and concepts were derived, but also to pinpoint the relevant text within the supporting document to make it easier for the user to find applicable facts and data in the response.

[0029] The system may include one or more devices, each device having at least one processor and a memory for storing the data and / or the executable code. The code may include software instructions for use in training the system as well as code configured to allow the system to interact with a user including an input and response capability using an input keypad, voice recognition, or another data input component or data interface, and one or more graphical user interface(s) (UI). In one aspect, the system may be programmed to include a plurality of Al agents for using the model to facilitate a legal or business transaction by an entity for exploiting its data. As used herein, references to Al agents' refer to the Al system or hybrid Al system described herein. In another aspect, the system may be coupled to, or may integrate with, a portal to an LLM such that the LLM may be employed to perform various ancillary actions or provide knowledge that furthers the ability of the entity to systematically and efficiently perform intended transactions, while leaving the core documentary concepts under the ambit of the specialized Al agents. The LLM’s ancillary role is strictly to augment, rather than substitute for, the precisely trained model used by the system to provide the necessary reliability when responding to user prompts. For example, in one embodiment, the LLM can be prompted by the user to provide summaries, translations or key points in the responses.

[0030] The above-referenced processor and memory systems may be organized into the one or more pre-trained, modifiable Al systems deployed to expedite the performance of functions relevant to the proposed licensing or other transaction involving a dataset. For example, the specialized Al system may provide knowledge and features highly relevant to the transaction(s) represented by forming the respective knowledge graph, which in turn is based on the training model or a modified version thereof. For example, In an embodiment, the processor-based Al platform for providing Software as a Service (SaaS) to a subscriber or licensee may be integrated with technology including the specialized Al system for performing the key tasks and for providing specialized responses to queries. The core concepts and domain information may be supplied by the processing resources in the cloud using the ground truth as implemented by the legal expert(s). The integration of the LLM for accessing areas in which it is superior with the informational substance of the prompt and query section to perform tasks governed by the ground truth beneficially allows the user to have access to both resources. For example, where the user wishes to modify the output of the application software in the cloud,the user can resort to LLM techniques to reformat the data. In another aspect of the disclosure, a main purpose of the LLM according to various embodiments is to (i) clarify the substance of the query that is being asked, (ii) confirm that the objects produced are relevant to that specific query. To avoid the statistical deficiencies of LLMs, the remainder of the process is dedicated to the pre-trained application software.

[0031] The existing solution makes dramatic improvements among prior solutions by integrating the capabilities of the domain knowledge (e.g., the set of knowledge graphs) and associated platform with the functionality of the specialized Al system and, In an embodiment, an LLM accessible by the user.

[0032] These and other aspects, objects, features, and advantages of the example embodiments will become apparent to those having ordinary skill in the art upon consideration of the following detailed description of example embodiments.BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Many aspects of the present disclosure can be better understood with reference to the following drawings. The components in the drawings are not necessarily to scale, emphasis instead being placed upon clearly illustrating the principles of the present disclosure. Moreover, in the drawings, like reference numerals designate corresponding parts throughout the several views.

[0001] An understanding of the features and advantages of the present invention will be obtained by reference to the following detailed description that sets forth illustrative embodiments, in which the principles of the invention may be utilized, and the accompanying drawings of which:

[0034] FIG. 1 depicts an example block diagram of a specialized Artificial Intelligence (Al)-enhanced system in which the principles of the present disclosure can be implemented.

[0035] FIG. 2 is an example of an excerpt including a prompt and a response from a traditional Retrieval-Augmented Generation (RAG) system using an LLM and a reasoning model illustrating the potential for inaccurate output produced by a probabilistic large language model-based retrieval system.

[0036] FIG. 3 is an example subset of a knowledge graph displayed on the user interface of a computing device comprising nodes representing documents, provisions, and domain concepts connected by named typed edges including read-for, related-to, supersedes, and similar relationships..

[0037] FIG. 4 is an example flow diagram of an approach to the specialized Al Agent / System using domain knowledge curated by domain experts to synthesize a response, in accordance with an embodiment.

[0038] FIG. 5A-B is an example display screen including a search indicating the different knowledge domains on a specialized Al system according to various embodiments, including an expanded menu for viewing exchange-based documents and FIG. 5B is an example expanded view of a document in the domain knowledge base displayed in response to a domain selection.

[0039] FIG. 6 is an example display screen providing a plurality of exchange-specific queries from which to choose and a Q&A section with a response based on the licensing requirements in an Alternative Trading System (ATS), wherein the plurality of exchangespecific queries are presented as a domain-specific query library of pre-built queries associated with the selected domain knowledge base, in accordance with an embodiment.

[0040] FIG. 7A-7B is an example display screen providing the response to the ATS query of FIG. 6, and in particular giving direct access to documentary sources supporting the response, with the key text highlighted with a boundary, together with the scores of different document sources and per-source relevance scores and source location identifiers for each documentary source cited in the response.

[0041] FIG. 8A-8E is an example display screen of a multi-document, multi-query processing matrix of the Al system, in which rows correspond to individual clients or counterparties, columns correspond to queries processed simultaneously across all documents in the domain knowledge base, and each cell contains a plain-language determination and extracted contractual language responsive to the corresponding query for the corresponding document, with a "No Available Response" indicator displayed in cells for which no responsive content is identified in the corresponding document. FIG. 8B illustrates a selected row state of the processing matrix corresponding to a client or counterparty document. FIG. 8C illustrates parsing of a document comprising multiple sub -documents, in which each sub-document is represented as a discrete cell entry within the corresponding matrix row. FIG. 8D illustrates a cell drill-down view of the processing matrix of FIG. 8 A, in which selection of a cell displays a source verification panel presenting the document hierarchy, section structure, and extracted source text relied upon to generate the cell response. FIG. 8E illustrates a column management interface providing sort and filter controls for individual query columns.

[0042] FIG. 9A-9G is an example display screen of a portfolio analytics dashboard of the Al system, comprising navigation tabs for accessing portfolio view modes including a ClientLiability Profile Snapshot, a Counterparty Profile Snapshot, a Counterparty vs. Issuer / Guarantor view, a Geographical Analysis view, and a Data Rights Image view; a nonconforming contracts table presenting contract metadata including contract type, country, contract term dates, and open exposure value for each client document in the corpus; an unexpired liability exposure panel presenting aggregate counts of identified liability provisions classified by liability category, with a filterable display control for selecting individual liability categories including confidentiality, gross negligence, indemnification obligation, IP infringement, privacy, and security breach; an expired liability analysis panel; and a client insurance coverage panel presenting aggregate risk status metrics across the document corpus.

[0043] FIG. 10 - A block flow diagram depicting methods to analyze structured domain documents and provide a plain-language response, in accordance with certain examples of the technology disclosed herein

[0044] Additional advantages of the disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or can be learned by practice of the disclosure. The advantages of the disclosure will be realized and attained by means of the elements and combinations particularly pointed out in the appended claims. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure, as claimed.

[0002] The figures herein are for illustrative purposes only and are not necessarily drawn to scale.DETAILED DESCRIPTION OF THE EXAMPLE EMBODIMENTS OVERVIEW

[0045] The various concepts introduced above and discussed in greater detail below may be implemented in any of numerous ways, as the described concepts are not limited to any particular manner of implementation. Examples of specific implementations and applications are provided primarily for illustrative purposes.

[0046] The present disclosure introduces techniques for using domain knowledge and a specialized Al system to answer accurately and free-formedquestions for users related to complex legal documents in financial services and other businesses. For example, a financial institution that has a subscription or license to the system may query the system to receive responses including domain data that reliably facilitates the creation of institution-relevant market data contracts with third parties and legal instruments related to data exploitation.

[0047] As noted in the Summary portion of this disclosure, companies need a clear understanding of the rights and obligations embedded in their contracts. While there are many contract types and use cases that benefit from the claimed technology, the inventors initially focus on the data contracts. However, determining rights and obligations remains a major problem. The principles of the disclosure make the rights and obligations available in an instant. For example, the principles of the disclosure can enable a computer, using a sophistical query / response architecture together with the visual and intuitive curating of the knowledge at issue, addresses this problem.System Architecture

[0048] Solution and key functionality. In an embodiment, the legal system in FIG. 1 (below) has at these and other functionalities in this area of interest: Analyzes and explains in plain English a company’s rights and obligations; Identifies conflicts and risks (e.g., ownership disputes, sublicensing restrictions); Compares amendments and historical versions to track contractual changes; Provides instant insights without reliance on a human being; Integrates with existing contract management systems via an application performing interface (API). As used herein, 'plain English' refers to multilingual and multimodal plain language.

[0049] In an embodiment of the disclosure, the claimed subject matter can be trained to manage complex legal documents with precision. The resulting savings in time and money are quantifiable. Built on legal expert-driven interpretive methods, it is not a black box, contextualizing, comprehending, interpreting, and delivering insights in plain English, not just excerpts that have been pattern matched and extracted. For more information on the process undertaken by the claims, please refer further to the Examples, including Example 1 and Example 2 at various places. Both appendices are incorporated by reference as if fully set forth herein.

[0050] FIG. 1 illustrates a specialized Al system 100 architecture. Workstation 102 serves as the user interface for financial institution agents or executives using the software application service (discussed below) in the cloud 114. A similar workstation or other computing device may be used by legal experts or their agents who upload training data in the manner described herein. This computing device contains processors (CPUs 104 and 106) connected to dynamic random access memory 108 via one or more integrated circuits for data routing. Workstation 102 may incorporate non-volatile storage (SSD 110) and may be implemented as a desktop, tablet, or other mobile device. The CPUs utilize various cache memory levels to optimize data access operations.

[0051] The workstation's processor(s) may employ RISC, CISC, or hybrid architectures. Alternative hardware implementations include digital signal processors (DSPs), applicationspecific integrated circuits (ASICs), systems-on-a-chip (SOCs), or field programmable, gate arrays (FPGAs). Solid state drive (SSD) 110 may store data and client code for accessing cloudbased applications. The executable software implements specific techniques described in this disclosure. User interaction occurs through UI 112 and various input peripherals. Network connectivity may be wireless or wired, with optional data encryption during transmission.

[0052] Workstation 102 connects to processing resources 124 via network 118 (e.g., the Internet, a virtual private network (VPN), wide area network (WAN), or cellular network). Cloud 114 may host application software executing on processing resources 124, which utilize high-capacity memory banks 122. The system may also include data repository 120 for storing legal expert training data that can be loaded into memory 122 for processing. An Al-enhanced large language model (LLM) 116 integrates with the network 118 to support specialized peripheral ancillary tasks for the Al system, such as clarifying the intent of and context for one or more queries prior to processing and confirming the relevance of returned objects following processing, and does not perform core document interpretation, knowledge graph traversal, or reasoning against the curated domain documents for Al functions. In some configurations, processing resources 124 execute server-based applications during communication with remote devices like workstation 102, the latter accessible by a user of the applications.

[0053] Alternative implementations may locate the application software locally on workstation 102. While system 100 represents one architectural example, variations are possible without departing from the disclosure's scope. The LLM may reside in separate cloud environments, and network 118 (shown within cloud 114 for simplicity) may be independently centralized. Further, in an embodiment, the method, system, and / or product may not involve or require using the cloud.

[0054] In an embodiment, the application code instructions described herein execute on a computing machine comprising one or more processors, one or more memory devices communicatively coupled to the one or more processors, one or more storage devices, and a network interface. In an embodiment, the one or more processors can comprise a general-purpose processor, a processor core, a multiprocessor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a tensor processing unit (TPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), a programmable logic device (PLD), or any combination thereof.

[0055] In an embodiment, the computing machine can include a graphics processing unit (GPU) or one or more co-processors. In an embodiment, a GPU comprises a specialized processing circuit having a highly parallel architecture that can accelerate processing of the application code instructions, including the processing operations described herein.

[0056] In an embodiment, the one or more storage devices can comprise a solid state drive (SSD), a hard disk drive, a compact disc (CD), a digital versatile disc (DVD), magnetic tape, flash memory, any other non-volatile memory device, any magnetic storage device, any optical storage device, any semiconductor storage device, or any combination thereof. A computer-readable storage device or medium, as used herein, does not include transitory signals, such as radio waves, electromagnetic waves propagating through a waveguide or other transmission media, or electrical signals transmitted through a wire.

[0057] In an embodiment, the network interface can facilitate communication between the computing machine and one or more other computing systems, servers, or networks. In an embodiment, such networks can comprise a local area network (LAN), a wide area network (WAN), a virtual private network (VPN), the Internet, a cellular network, or any combination thereof.

[0058] In an embodiment, the methods described herein can be implemented as computerexecutable program instructions embodied on a non-transitory computer-readable storage device that, when executed by a computer, cause the computer to perform the operations described herein. In an embodiment, such computer-executable program instructions can be executed by one or more processors of the computing machine described herein, or by one or more processors of a distributed computing environment communicatively coupled via a network.

[0059] In an embodiment, the techniques described herein can be implemented as a system, a method, a computer program product, digital electronic circuitry, computer hardware, firmware, software, or any combination thereof.

[0060] In an embodiment, the computer-executable program instructions can comprise assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, source code, object code, or any combination thereof.

[0061] In an embodiment, the computer-executable program instructions can be downloaded from a remote computer-readable storage medium or device to a local computer-readable storage medium or device for storage and subsequent execution by a computer. In anembodiment, the computer-executable program instructions can be transmitted via one or more networks for storage on and execution by a remote computer-readable storage device.

[0062] In an embodiment, the Al system described herein is implemented as a specialized small language model (SLM). An SLM, as used in this disclosure, is a language model trained on a bounded, curated domain knowledge base that is specific to a defined document library. The processing of the SLM can be configured to be deterministic and rule-based, drawing exclusively from structured domain documents assembled and verified by human domain experts during the knowledge curation phase described herein.

[0063] In an embodiment, the system can incorporate an LLM as an ancillary module to perform peripheral functions, including without limitation clarifying the intent of one or more queries prior to processing and confirming the relevance of returned objects following processing; in such embodiments, core document interpretation, traversal of the domain knowledge base, and reasoning against the curated domain documents are performed by the deterministic, knowledge-graph-based processing architecture. The term "Al system," as used in this disclosure, encompasses a small language model, a large language model, a latent Dirichlet allocation model, a neural network, or a combination thereof configured to perform the processing functions described herein.Trainins Methods Overview

[0064] In an embodiment, the Al system can be trained using techniques such as unsupervised learning, supervised learning, transfer learning, incremental learning, and / or curriculum learning. In general, training occurs after selection and development of an Al system and before the Al system is operably in use. In an embodiment, the training data used to teach the Al system can comprise input data such as structured domain documents and the respective target output data such as plain-language responses, source references, domain knowledge classifications, or a combination thereof.

[0065] In an embodiment, training of the Al system can comprise between about 1 and about 100 training epochs, for example between about 5 and about 50 epochs, with a batch size of between about 1 and about 10,000 structured domain documents per training step.Unsupervised and Supervised Learning

[0066] In an embodiment, unsupervised learning is implemented. Unsupervised learning can involve providing all or a portion of unlabeled training data to an Al system. The Al system can then determine one or more outputs implicitly based on the provided unlabeled training data. In an embodiment, supervised learning is implemented. Supervised learning can involve providing all or a portion of labeled training data to an Al system, with the Al systemdetermining one or more outputs based on the provided labeled training data, and the outputs are either accepted or corrected depending on the agreement to the actual outcome of the training data. In some examples, supervised learning of an Al system can be governed by a set of rules and / or a set of labels for the training input, and the set of rules and / or set of labels can be used to correct inferences of the Al system.Transfer Learning

[0067] In an embodiment, transfer learning is implemented. Transfer learning techniques can involve providing all or a portion of a first training data to an Al system, then, after training on the first training data, providing all or a portion of a second training data. In an embodiment, a first Al system can be pre-trained on data from one or more computing devices. The first trained Al system is then provided to a computing device, where the computing device is intended to execute the first trained Al system to produce an output. Then, during the second training phase, the first trained Al system can be additionally trained using additional training data, where the training data can be derived from structured domain documents associated with one or more computing devices. This second training of the Al system using the training data can be performed using either supervised, unsupervised, or semi-supervised learning. In addition, it is understood that transfer learning techniques can involve one, two, three, or more training attempts. Once the Al system has been trained on at least the training data, the training phase can be completed. The resulting trained Al system can be utilized as at least one trained Al system.Incremental and Curriculum Learning

[0068] In an embodiment, incremental learning is implemented. Incremental learning techniques can involve providing a trained Al system with input data that is used to continuously extend the knowledge of the trained Al system. Another training technique is curriculum learning, which can involve training the Al system with training data arranged in a particular order, such as providing training examples first, then proceeding with progressively more complex training examples.Pre-Trained Learning

[0069] In an embodiment, the Al system is pre-trained. A pre-trained Al system is an Al system that has been previously trained to address a related problem. The pre-trained Al system is generally pre-trained with input data similar to that of the new application. A pre-trained Al system further trained to address a new problem is a technique referred to as transfer learning. In some instances, a pre-trained Al system is trained on a large dataset of related information. The pre-trained Al system is then further trained and tuned for the new application. Using apre-trained Al system provides the advantage of building a new Al system with input neurons / nodes that are already familiar with the input data and are more readily refined to a particular problem. For example, an Al system previously trained using structured domain documents can be further trained to generate plain-language responses, source references, domain knowledge classifications, or a combination thereof.

[0070] In some examples, after the training phase has been completed but before producing predictions expressed as outputs, a trained Al system can be provided to a computing device where a trained Al system is not already resident; in other words, after the training phase has been completed, the trained Al system can be downloaded to a computing device. For example, a first computing device storing a trained Al system can provide the trained Al system to a second computing device. Providing the trained Al system to the second computing device can comprise one or more of communicating a copy of the trained Al system to the second computing device, making a copy of the trained Al system for the second computing device, providing access to the trained Al system to the second computing device, and / or otherwise providing the trained Al system to the second computing device. In an embodiment, a trained Al system can be used by the second computing device immediately after being provided by the first computing device. In some examples, after a trained Al system is provided to the second computing device, the trained Al system can be installed and / or otherwise prepared for use before the trained Al system can be used by the second computing device.

[0071] After an Al system has been trained, it can be used to output, estimate, infer, predict, generate, produce, or determine results, and these terms are collectively referred to as results. A trained Al system can receive input data and operably generate results. As such, input data can be used as an input to the trained Al system for providing corresponding results to downstream processing components. For example, a trained Al system can generate results in response to requests. In an embodiment, a trained Al system can be executed by a portion of other software. For example, a trained Al system can be executed by a result daemon to be readily available to provide results upon request.

[0072] In an embodiment, an Al system and / or trained Al system can be executed and / or accelerated using one or more computer processors and / or on-device co-processors. Such on-device co-processors can speed up training of an Al system and / or generation of results. In some examples, a trained Al system can be trained, reside, and execute to provide results on a particular computing device, and / or otherwise can make results available for the particular computing device.

[0073] Input data can include data from a computing device executing a trained Al system and / or input data from one or more computing devices. In an embodiment, a trained Al system can use results as input feedback. A trained Al system can also rely on past results as inputs for generating new results. In an embodiment, input data can comprise one or more queries, or structured domain documents, and, when provided to a trained Al system, results in output data such as plain-language responses, source references, domain knowledge classifications, or a combination thereof. The output can then be provided to downstream workflow processing. The downstream workflow processing can use the output to inform the user's analytical, legal, or financial operations, thereby providing a source-attributed response to the user's query. As such, the query-processing-related problems addressed by this disclosure can be resolved using the herein-described techniques that utilize the Al system to produce plain-language responses, source references, or domain knowledge classifications for use in downstream workflow processing.Machine Learning ModelsSmall Language Model

[0074] In an embodiment, the Al system is implemented as a small language model (SLM). An SLM is a language model trained on a bounded, curated domain-specific corpus. The SLM's learned representations are derived from structured domain documents contributed by or on behalf of the user, including legal agreements, regulatory filings, compliance policies, and related domain materials. In an embodiment, the SLM can process one or more queries against a curated domain knowledge base and generate plain-language responses, source references, domain knowledge classifications, or a combination thereof. In an embodiment, the SLM can produce a consistent output for a given input query and domain knowledge base configuration, such that the same query applied to the same domain knowledge base can produce the same plain-language response and source references. In an embodiment, the SLM can be deployed as a dedicated Al system associated with a particular user's document library, with the SLM's learned representations stored in a dedicated partition associated with the user and inaccessible to other users or Al systems associated with other users.Large Language Model

[0075] In an embodiment, the Al system can incorporate a large language model (LLM) as an ancillary module. An LLM is a language model trained on a large, general -purpose corpus comprising diverse text sources. LLMs can generate human -readable text in response to natural language prompts. In an embodiment, an LLM can perform peripheral functions that augment the processing performed by the Al system against the domain knowledge base, with coredocument interpretation, knowledge graph traversal, and reasoning performed by the Al system. For example, prior to the Al system processing a user query, an LLM can receive the user's query and return a clarified or rephrased formulation of the query to the Al system for downstream processing against the domain knowledge base. In an embodiment, following retrieval of a plain-language response and source references from the domain knowledge base, an LLM can evaluate the relevance of the returned response to the original query and provide a relevance confirmation to the user. In an embodiment, the LLM operates at the query input and response output stages only.Neural Networks

[0076] In an embodiment, neural networks (NNs) are implemented. NNs are a family of statistical learning models influenced by biological neural networks of the brain. NNs can be trained on a relatively large dataset (e.g., 50,000 or more) and used to estimate, approximate, or predict an output that depends on a large number of inputs / features. NNs can be envisioned as "neuromorphic" systems of interconnected processor elements, or "neurons," that exchange electronic signals, or "messages." Similar to the "plasticity" of synaptic neurotransmitter connections that carry messages between biological neurons, the connections in NNs that carry electronic "messages" between "neurons" are provided with numeric weights that correspond to the strength or weakness of a given connection. The weights can be tuned based on experience, making NNs adaptive to inputs and capable of learning. For example, an NN for generating plain-language responses, source references, or domain knowledge classifications is defined by a set of input neurons that can be given input data such as one or more queries, or structured domain documents. The input neuron weighs and transforms the input data and passes the result to other neurons, often referred to as "hidden" neurons. This is repeated until an output neuron is activated. The activated output neuron produces a result. In an embodiment, one or more queries, or structured domain documents, are used to train the neurons in a NN Al system, which, after training, is used to estimate plain-language responses, source references, domain knowledge classifications, or a combination thereof.

[0077] In an embodiment, a NN layer (e.g., input, hidden, and output) has 1 or more neurons. In an embodiment, the NN layer has 1-500 neurons. In an embodiment, the NN layer has about 1 neuron, 5 neurons, 10 neurons, 15 neurons, 20 neurons, 25 neurons, 30 neurons, 35 neurons, 40 neurons, 45 neurons, 50 neurons, 55 neurons, 60 neurons, 65 neurons, 70 neurons, 75 neurons, 80 neurons, 85 neurons, 90 neurons, 95 neurons, 100 neurons, 105 neurons, 110 neurons, 115 neurons, 120 neurons, 125 neurons, 130 neurons, 135 neurons, 140 neurons, 145 neurons, 150 neurons, 155 neurons, 160 neurons, 165 neurons, 170 neurons, 175neurons, 180 neurons, 185 neurons, 190 neurons, 195 neurons, 200 neurons, 205 neurons, 210 neurons, 215 neurons, 220 neurons, 225 neurons, 230 neurons, 235 neurons, 240 neurons, 245 neurons, 250 neurons, 255 neurons, 260 neurons, 265 neurons, 270 neurons, 275 neurons, 280 neurons, 285 neurons, 290 neurons, 295 neurons, 300 neurons, 305 neurons, 310 neurons, 315 neurons, 320 neurons, 325 neurons, 330 neurons, 335 neurons, 340 neurons, 345 neurons, 350 neurons, 355 neurons, 360 neurons, 365 neurons, 370 neurons, 375 neurons, 380 neurons, 385 neurons, 390 neurons, 395 neurons, 400 neurons, 405 neurons, 410 neurons, 415 neurons, 420 neurons, 425 neurons, 430 neurons, 435 neurons, 440 neurons, 445 neurons, 450 neurons, 455 neurons, 460 neurons, 465 neurons, 470 neurons, 475 neurons, 480 neurons, 485 neurons, 490 neurons, 495 neurons, 500 neurons, or any range between any two number of neurons listed.Deep Learning

[0078] In an embodiment, deep learning is implemented. Deep learning expands the neural network by including more layers of neurons. A deep learning module is characterized as having three "macro" layers: (1) an input layer which takes in the input features, and fetches embeddings for the input, (2) one or more hidden layers which process the input, and (3) an output layer which provides the result.Topic Models

[0079] In an embodiment, a topic model can refer to a class of probabilistic machine learning models used for text classification and latent topic inference, wherein a collection of documents is analyzed to identify abstract thematic structures underlying the text. Topic models can include, for example, latent Dirichlet allocation (LDA), non-negative matrix factorization (NMF), probabilistic latent semantic analysis (pLSA), correlated topic models, hierarchical topic models, or other statistical models that represent documents as mixtures over a set of latent topics.Latent Dirichlet Allocation (LDA)

[0080] In an embodiment, one or more queries, or structured domain documents, are processed with Latent Dirichlet Allocation (LDA). LDA is a topic model for classifying text, wherein a document or more generally a set of text represents a random mixture over latent topics and each topic is characterized by a distribution of words. LDA is capable of identifying similar groups of text and associating them with certain topics. Generally, topics are identified by searching for groups of text in a document and taking a probability distribution that a group of text belongs to a topic and is likely to be found in the document.

[0081] In an embodiment, LDA is used on structured domain documents to identify domain knowledge classifications, plain-language responses, source references, or a combinationthereof. First, a number of topics are selected to be determined from the structured domain documents. The topics can comprise any topic related to the domain knowledge classifications described herein. The LDA model is then trained to learn the selected topics. A set of training text is used as input for the LDA model. The text is randomly distributed among the selected topics. In an iterative process, the LDA model determines the proportion of text in a set that are currently assigned to a selected topic, then determines the proportion of assignments to the selected topic over all the sets and reassigns the word to a different topic based on a computed probability. This process is complete once a steady state of acceptable assignments is determined. The LDA model can then be used to determine topics from user queries, which can then be passed to the Al system. See e.g., Blei, David M., Andrew Y. Ng, and Michael I. Jordan. "Latent dirichlet allocation." Journal of machine Learning research 3. Jan (2003): 993- 1022, incorporated herein by reference.Natural Language Processing

[0082] In an embodiment, one or more queries, or structured domain documents, are processed with a natural language processing (NLP) model. NLP is a computational approach to evaluating text. General applications of NLP can comprise information retrieval, information extraction, question-answering, summarization, machine translation, and dialogue systems. Information extraction comprises recognition, tagging, and extraction into a structured representation. For example, key elements of information, such as persons, companies, locations, and / or organizations, are extracted from collections of text. The key elements of information can include physical properties (e.g., quantifiable / measurable) recited in the text and / or subjective properties (e.g., interpretive assessments). Question-answering (QA) comprises generating a list of documents corresponding to a user's query. For example, QA can comprise generating either the text of the answer or answer-providing passages. Summarization comprises reducing a large amount of text into a smaller amount of text. For example, summarization can comprise an abbreviated narrative representation of the original document. Machine translation (MT) comprises either rule-based or probabilistic methods (e.g., machine learning) for translating text to speech or vice versa, or from one language to another. For example, MT attempts to capture contextual, idiomatic, and pragmatic aspects of language. Dialogue systems comprise conversational communications through communication modes such as text, speech, and images.

[0083] NLP can comprise different levels of language. For example, ascending levels of language can comprise phonology, morphology, lexical, syntactic, semantic, discourse, and pragmatic. The phonology level comprises the interpretation of speech sounds. For example,phonological analysis can comprise phonetic rules, phonemic rules, and prosodic rules. Morphology comprises the nature of words, which comprise morphemes (i.e., smallest units of meaning). For example, the suffix -ed added to a verb indicates the action of the verb took place in the past. The lexical level comprises interpreting the meaning of individual words. For example, words can be assigned part-of-speech tags based on context. The syntactic level comprises analyzing the words in a sentence to extract the grammatical structure of the sentence. For example, syntactic processes attempt to compute the meaning of a sentence from the order and dependency of the words in a sentence. The semantic level comprises determining the meaning of a sentence by analyzing the interactions between word-level meanings in a sentence. For example, a semantic process can comprise disambiguation wherein words with multiple meanings are reduced to a singular meaning based on the context of the sentence. The discourse level comprises determining the meaning of more than one sentence. For example, discourse processing can comprise anaphora resolution and discourse / text structure recognition. The pragmatic level comprises determining the use of language based on the context of the text, not necessarily the content of the text. For example, pragmatic processes deduce extra meanings read into text that are not necessarily encoded in the text.

[0084] NLP can also comprise different approaches to language processing. For example, the approaches can comprise symbolic, statistical, connect! oni st, and hybrid approaches. Symbolic approaches comprise using explicit representations of facts through well -understood knowledge representation schemes to analyze linguistic phenomena. For example, symbolic approaches can comprise logic or rule-based systems. Statistical approaches comprise using text corpora to determine generalized models of linguistic phenomena. For example, a statistical approach can use a hidden Markov model (HMM) to determine speech recognition, lexical acquisition, parsing, part-of-speech tagging, collocations, statistical machine translation, statistical grammar learning, and similar tasks. Connectionist approaches comprise combining statistical learning with various theories of representation. For example, a connectionist approach can comprise a network of interconnected local processing units with knowledge stored as weights in the connections between units, wherein local interactions can result in an observed global behavior. See e.g., Liddy, E.D. 2001. Natural Language Processing. In Encyclopedia of Library and Information Science, 2nd Ed. NY. Marcel Decker, Inc.

[0085] NLP tasks can comprise: sentence boundary detection (e.g., abbreviations and titles - "e.g.," "Dr."); tokenization (e.g., hyphens, forward slashes - "§3.5," "cross-reference"); part-of-speech assignment to individual words (e.g., homographs and gerunds); morphologicaldecomposition (e.g., lemmatization); shallow parsing (chunking) (e.g., identifying phrases); problem-specific segmentation (e.g., segmenting text into meaningful groups); spelling / grammatical error identification and recovery (e.g., recovering false positives); named entity recognition (e.g., identifying persons, locations, organizations, or regulatory bodies); word sense disambiguation (e.g., determining a homograph's correct meaning in a given domain context); negation and uncertainty identification (e.g., inferring whether a named entity or obligation is present or absent); relationship extraction (e.g., determining relationships between entities or provisions); temporal inferences / relationship extraction (e.g., inferring that something has occurred in the past or can occur in the future); and information extraction (e.g., extracting relevant provisions from a document involves named entity recognition, word sense disambiguation, negation detection, temporal inference, and anaphoric resolution).Bayesian Network

[0086] In an embodiment, the domain knowledge base can be represented as or processed using a Bayesian network. A Bayesian network is a probabilistic graphical model in which nodes represent domain variables and directed edges encode conditional dependencies between those variables. In an embodiment, the knowledge graph of the domain knowledge base can be structured as a Bayesian network in which document provisions, statutory references, and compliance requirements form the nodes, and the conditional relationships among those provisions are encoded as directed edges. The Al system can use the structure of the Bayesian network to identify which document provisions are conditionally relevant to a user query given the relationships established during the training phase.Decision Trees

[0087] In an embodiment, the domain knowledge base can represent domain decision logic as one or more decision trees. A decision tree is a branching graph comprising nodes, edges, and leaf nodes. Each internal node represents a decision point based on a domain criterion, and each edge represents an outcome of that decision, directing traversal toward a subsequent node or leaf. Leaf nodes represent terminal determinations, such as the applicability of a document provision, a compliance classification, or a jurisdiction-specific rule. In an embodiment, the Al system can traverse a decision tree during query processing to determine which provisions of the domain knowledge base are responsive to the user's query and to apply domain-specific priority rules among those provisions.Frame-Based Systems

[0088] In an embodiment, the domain knowledge base can be organized using a framebased representation. A frame-based system organizes domain knowledge into structured unitscalled frames, each representing a domain concept, document type, or provision category. Each frame comprises a set of slots that define the attributes, relationships, and default values associated with the frame. In an embodiment, frames representing document provisions can include slots for the applicable jurisdiction, the document type, associated cross-references, and the priority hierarchy of the provision relative to other provisions in the domain knowledge base. The Al system can retrieve relevant frames during query processing to identify the set of document provisions and relationships responsive to the user query.Conceptual Graphs

[0089] In an embodiment, the domain knowledge base can be represented as one or more conceptual graphs. A conceptual graph is a formal knowledge representation in which concept nodes represent domain entities, events, or states, and relation nodes represent the relationships among those concepts. In an embodiment, the domain knowledge base can be represented as a conceptual graph in which concept nodes correspond to document provisions, statutory terms, or domain-specific entities, and relation nodes encode the semantic relationships among those provisions established during the training phase. The Al system can traverse the conceptual graph during processing to identify the sequence of domain relationships responsive to a user query.Description Logic Knowledge Bases

[0090] In an embodiment, the domain knowledge base can be structured as a description logic knowledge base. Description logic is a family of formal knowledge representation languages for defining domain concepts, roles, and individuals using logical axioms. A description logic knowledge base comprises a terminological component defining the domain concept hierarchy and role relationships, and an assertional component specifying the properties of individual domain entities. In an embodiment, the terminological component can define the hierarchy of document types, regulatory frameworks, and compliance categories relevant to the domain, while the assertional component can specify the properties and relationships of individual document provisions, statutory sections, and contract clauses within the user's document library. The Al system can use the description logic knowledge base during query processing to identify document provisions that satisfy the logical conditions expressed by the user's query.Embedding

[0091] In an embodiment, the Al system can use embedding to provide a lower dimensional representation, such as a vector, of features to organize them based on respective similarities. In some situations, these vectors can become large. In the case of large vectors,particular values can become sparse among a large number of values (e.g., a single instance of a value among 50,000 values). Because such vectors are difficult to work with, reducing the size of the vectors can be advantageous. An Al system can learn the embeddings along with the model parameters. In an embodiment, features such as one or more queries, or structured domain documents, can be mapped to vectors implemented in embedding methods. In an embodiment, embedded semantic meanings are utilized. Embedded semantic meanings are values of respective similarity. For example, the distance between two vectors, in vector space, can imply that two values located elsewhere with the same distance are categorically similar. Embedded semantic meanings can be used with similarity analysis to rapidly return similar values. In an embodiment, one or more queries, or structured domain documents, are embedded. In an embodiment, the methods described herein are developed to identify meaningful portions of the vector and extract semantic meanings between that space.Floating-point Operations Per Second (FLOPS)

[0092] FLOPS are a measure of computation speed. The base level for computing is gigaFLOPS, which is 1 billion (10A9) floating-point operations per second. This would be equivalent to a person with pen and paper or a calculator constantly performing one (1) calculation every one (1) second for approximately thirty-two (32) years. As computing floating-point numbers is necessary in fields such as financial applications, scientific applications, visual rendering, and real-time processing, this unit of measurement can be used to describe the methods, systems, and products described herein. As such, and further described herein, the methods, systems, and products are carried out on time scales larger than a couple of seconds. Accordingly, these methods, systems, and products are required to be carried out computationally because no person in a lifetime could carry out these methods.

[0093] To illustrate the disparity between human and computer capabilities, consider the simple task of sorting a list of one million numbers. A modern computer can complete this in approximately 100 milliseconds. In contrast, a human attempting this task manually, even under unrealistically optimal conditions, would require about 11.5 days of continuous work. Realistically, accounting for fatigue and necessary breaks, this task could take weeks or months for a human to complete. The methods, systems, and products described herein are significantly more complex than this simple sorting example. They require an amount of time that is, for all practical purposes, near infinite for a human to complete manually. The computational complexity and scale of the task place it in the realm of processes that can only be executed by purpose-built computing systems, not by the human mind.Cloud and Software as a Service

[0094] In an embodiment, the present techniques can be at least partially implemented in a cloud or virtual machine environment. One or more processors can also operate to support performance of the relevant operations in a "cloud computing" environment or as a "software as a service" (SaaS). For example, at least some of the operations can be performed by a group of computers, these operations being accessible via a network and via one or more appropriate interfaces (e.g., application program interfaces (APIs)).Service Models

[0095] Software as a Service (SaaS): the capability provided to the consumer is to use the provider's applications running on a cloud infrastructure. The applications are accessible from various client devices through a thin client interface such as a web browser (e.g., web-based e-mail). The consumer does not manage or control the underlying cloud infrastructure including network, servers, operating systems, storage, or even individual application capabilities, with the possible exception of limited user-specific application configuration settings.

[0096] Platform as a Service (PaaS): the capability provided to the consumer is to deploy onto the cloud infrastructure consumer-created or acquired applications created using programming languages and tools supported by the provider. The consumer does not manage or control the underlying cloud infrastructure including networks, servers, operating systems, or storage, but has control over the deployed applications and possibly application hosting environment configurations.

[0097] Infrastructure as a Service (laaS): the capability provided to the consumer is to provision processing, storage, networks, and other fundamental computing resources where the consumer is able to deploy and run arbitrary software, which can include operating systems and applications. The consumer does not manage or control the underlying cloud infrastructure but has control over operating systems, storage, deployed applications, and possibly limited control of select networking components (e.g., host firewalls).

[0098] A cloud computing environment is service oriented with a focus on statelessness, low coupling, modularity, and semantic interoperability. At the heart of cloud computing is an infrastructure comprising a network of interconnected nodes.Cloud Computing Environment

[0099] A cloud computing environment can include one or more cloud computing nodes with which local computing devices used by cloud consumers, such as, for example, personal digital assistants (PDAs) or cellular telephones, desktop computers, laptop computers, and / or automobile computer systems can communicate. The nodes can communicate with oneanother. The nodes can be grouped physically or virtually, in one or more networks, such as Private, Community, Public, or Hybrid clouds as described herein, or a combination thereof. This allows the cloud computing environment to offer infrastructure, platforms, and / or software as services for which a cloud consumer does not need to maintain resources on a local computing device. It is understood that the types of computing devices are intended to be illustrative only and that computing nodes and the cloud computing environment can communicate with any type of computerized device over any type of network and / or network-addressable connection (e.g., using a web browser).

[0100] Hardware and software layer includes hardware and software components. Examples of hardware components include mainframes; RISC (Reduced Instruction Set Computer) architecture-based servers; servers; blade servers; storage devices; and networks and networking components. In an embodiment, software components include network application server software and database software.

[0101] Virtualization layer provides an abstraction layer from which the following examples of virtual entities can be provided: virtual servers; virtual storage; virtual networks, including virtual private networks; virtual applications and operating systems; and virtual clients.

[0102] In an embodiment, the method, system, product, and / or elements thereof, can be implemented in the form of an apparatus including a memory and at least one processor that is coupled to the memory and operative to perform exemplary method steps. A computer system can be useful in implementing one or more aspects and / or elements described herein. The cloud computing node is only one example of a suitable cloud computing node and is not intended to suggest any limitation as to the scope of use or functionality of embodiments as described herein. Regardless, the cloud computing node is capable of being implemented and / or performing any of the functionality set forth hereinabove.

[0103] In a cloud computing node, there is a computer system / server that is operational with numerous other general purpose or special purpose computing system environments or configurations. Examples of well-known computing systems, environments, and / or configurations that can be suitable for use with the computer system / server include, but are not limited to, personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments that include any of the above systems or devices.

[0104] The computer system / server can be described in the general context of computer system executable instructions, such as program modules, being executed by a computer system. Generally, program modules can include routines, programs, objects, components, logic, data structures, and so on that perform particular tasks or implement particular abstract data types. The computer system / server can be practiced in distributed cloud computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed cloud computing environment, program modules can be located in both local and remote computer system storage media including memory storage devices.Deployment Models

[0105] Private cloud: the cloud infrastructure is operated solely for an organization. It can be managed by the organization or a third party and can exist on-premises or off-premises.

[0106] Community cloud: the cloud infrastructure is shared by several organizations and supports a specific community that has shared concerns (e.g., mission, security requirements, policy, and compliance considerations). It can be managed by the organizations or a third party and can exist on-premises or off-premises.

[0107] Public cloud: the cloud infrastructure is made available to the general public or a large industry group and is owned by an organization selling cloud services.

[0108] Hybrid cloud: the cloud infrastructure is a composition of two or more clouds (private, community, or public) that remain unique entities but are bound together by standardized or proprietary technology that enables data and application portability (e.g., cloud bursting for load-balancing between clouds).

[0109] In an embodiment, in forming a response to a query, the Al system draws exclusively from the curated domain knowledge base and the associated document library. No external network, internet-based data source, or data repository outside the curated domain knowledge base is accessed during the response formation process. Because the system forms responses only from curated domain documents, each response can be traceable to identified source passages within the user's document library.Domain Knowledge Base

[0110] Using a market data contract as an example of a contract type curated by the invention, a market data contract is a complex legal agreement to aquire a license to to data provided by a licensor that governs the buying, selling, licensing, or exchange of data between parties. In general, a market data contract may be entered into by a data provider (typically a financial exchange, market data vendor, or financial information provider) and an end user(such as a financial institution, trader, or investment firm) that governs the terms and conditions for accessing, using, and distributing financial market data. Market data contracts can involve copious terms and conditions and are often accompanied by one or more schedules, policies, exhibits, appendices, etc. These contracts typically cover several key aspects of the data-based transaction, including, for example, the scope of the data being provided (real-time quotes, historical data, etc.), permitted usage rights and restrictions, e-distribution limitations, pricing structure and payment terms, compliance requirements, audit rights for the data provider, intellectual property protections, confidentiality clauses, and term length and termination conditions, among others.[OHl] A major drawback that an entity affiliated with the user may face is how to reliably interpret data exploitation objects such as market data contracts, the latter of which as noted may include complex nuances, a dedicated ontology, and a body of knowledge specific to a fact situation or anticipated legal transaction. Referring to any of the many data exchanges such as the Chicago Mercantile Exchange (CME), used as an example herein, these exchanges may utilize two dozen or more such contracts, each with their own specific objective depending on the nature of the transaction. Market data contracts give rise to a number of initial questions for the entity. Non-exhaustive examples of these questions from a financial institution or other entity may include, for example, whether the institution owns the data, whether it can reproduce the data, whether or not it can create derivative works based on the data, whether it can sell products based on the derived data, etc. Due to their complexity and the levels of precision involved in such agreements, the relevant documents and principles may be overwhelming to non-lawyers. For example, in lieu of spending days or weeks interpreting a single fully executed or weeks examining multiple agreements, a user (e.g., an attorney) may employ an Al-enhanced system for inputting queries that produce responses to help guide the user through the factual and legal hurdles of the agreement, and to respond to changing marketplace dynamics in minutes. Recent LLM breakthroughs have resulted in tantalizing new frontiers in legal automation, yet legal documents remain a strategic challenge for at least the following reasons:

[0112] Complexity and Nuance: Legal texts often contain intricate clauses, context-dependent terminology, and ambiguous language. High-stakes legal documents like market data contracts feature extensive definitions and cross-references requiring a comprehensive understanding of their structure.

[0113] Accuracy Demands: In the legal arena, small errors can have major consequences. Misinterpreting ownership clauses in a market data agreement, for example, may expose organizations to significant financial and regulatory risks.

[0114] Dynamic Nature: Laws and legal interpretations evolve continuously. Al systems must adapt their knowledge to remain relevant and accurate guides for this information.

[0115] Conventional Retrieval-Augmented Generation (RAG) systems — systems that combine traditional document retrieval techniques with LLM-generated responses — stumble when navigating these subtleties, spawning errors that undermine their utility. FIG. 2 shows an example of a RAG-based dialogue excerpt including a prompt and a response from an LLM using a detailed reasoning model (here, armed with 20 PDF documents). The model confidently misdefines "original works" and delivers a dangerously incorrect answer with unwavering conviction.

[0116] Existing RAG systems include additional deficiencies. For example, RAG-based systems often limit their searches to rely on keyword matching and statistical relevance. This limited search mechanism can miss the presence of deeper semantic relationships, implied conditions, and other subtleties. This deficiency is particularly noteworthy in areas such as law, where the subtleties and context in an excerpt of text, such as a statutory provision referenced in an agreement, are often extremely important. In another instance, ontology mismatches can impair the search results. In the context of Al, ontology refers to a conceptualization that represents knowledge as a set of concepts within a domain and the relationships between those concepts. Different bodies of legal documents use their own unique ontologies. Absent a structured approach to the search process, reconciling differing terminologies and definitions across multiple sources becomes challenging.

[0117] As still another example, conventional RAG-based systems often output data that results in a loss of structural complexity. Legal documents typically have hierarchical clause structures, cross-references, and dedicated definition sections. Traditional RAG pipelines are not designed to preserve or interpret this level of detail, which can result in an incomplete or incorrect analysis. Further, as noted above, LLMs are capable of generating plausible but factually incorrect interpretations of data if not guided by precise and validated context. Hallucinations are particularly problematic in high-stakes legal settings, where precision is often imperative.

[0118] In an embodiment, the document library is stored in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users, such that documents contributed by or on behalf of one user cannot beaccessed by or used in processing queries associated with any other user. In an embodiment, the per-user dedicated partition can support confidentiality of the document library in environments subject to data privacy, cybersecurity, or professional confidentiality requirements. In an embodiment, the document library is also organized in said dedicated partition.The Training Method

[0119] To overcome these and other deficiencies with conventional RAG systems in one aspect of the disclosure, knowledge curation is used. Knowledge curation is a deliberate, expert-driven process that systematically captures, organizes, and validates domain-specific information before integrating it into Al systems. Among other benefits, using knowledge curation ensures that the Al system has reliable, up-to-date knowledge, mitigating many of the pitfalls observed in traditional RAG training techniques.

[0120] In an embodiment, the specialized Al application includes an initial instance of the application layer of the system. The present instance or data (e.g., a response to a query, knowledge graph, reference data, etc.) may be displayed to a user of the system on one or more UIs. As an illustration, the initial display screen may include a region identifying the distributor of the application, along with text providing an initial greeting like: “Welcome! Let’s get to know you”. These headings may be followed by input icons requesting the user’s name. It is noteworthy that this example is one of many, and the initial UI may have different selections or information in other implementations. This example may represent the starting point of the system when the user signs on. The initial screen like the one described above may act as a figurative springboard upon which the user or financial entity can begin to access more specific data tailored to answer the types of questions presented and, where applicable, to provide the identity of specific documents. The initial actions by the user may also coincide with the commencement of knowledge curation specific to a query.

[0121] The present Al system allows the entity to circumvent the limitations described above with respect to LLMs by providing a specialized Al system having data whose purpose is to reliably questions regarding an exchange or other entity. The intuitive knowledge base provided by the present system can enable the entity to enter into a transaction fully formed, to consider a transaction proposed by a third party, or simply to educate itself with respect to the current state of the entity’s interests in, and entitlement to, the exploitation of data.

[0122] In an embodiment, display, the Al agents described herein may initially present a dialogue box, such as the one shown “Q&A” to enable the user to ask a question of any level of difficulty to facilitate an entity’s understanding or intent to use the data as part of aprospective transaction to collect revenue. For example, the user may proceed to open a dialogue box from the application, which in this example provides an entry tool for a custom query. The user may enter a query of any level of complexity regarding the selected exchange or other entity. The query may be anything from an isolated request for information to one of a successive number of queries for entering into revenue-generating transactions with third parties. In an embodiment, the user may ask a straightforward question by entering the following input in the query box: “What is the cut off for delayed data?”. Delayed data refers to information or datasets that are not immediately available or updated in real-time but instead are received or processed after a certain period. This delay may be due to various factors, such as the time required for data collection, transmission, or processing, and the pre-training time period. For example, with reference to different data exchanges, the data may by necessity be reported with a time lag (such as economic statistics or stock prices). As discussed in connection with FIG. 5 (below), the user may identify the query as relevant..

[0123] Knowledge curation has the following exemplary characteristics. For one, knowledge curation enables the system to identify critical concepts and clauses, such as in a market data agreement. Experts collaborate with automated tools to pinpoint key definitions, relevant clauses, and domain-specific terminologies. In addition, knowledge curation includes cross-checks of extracted information with legal experts such as senior attorneys to ensure accuracy. This characteristic is particularly valuable for market data agreements, where even subtle differences in wording can drastically change interpretations of pivotal legal concepts such as data ownership or usage rights. Yet another benefit of knowledge curation is that word definitions and relationships between words or terms can be standardized, such as by being normalized and organized into a consistent ontology. This characteristic is essential for the Al system to reconcile different document sources as well as any overarching cross-exchange framework. Another well-known fact is that legal rules and regulations often evolve over time. Typical LLMs may employ an arbitrary training “cut off’ where information after the cut off (such as repealed statutes in the legal context) is unavailable to the user. To more closely accommodate changes in the relevant laws or other information and thereby reduce the risk of outdated interpretations, the system may undergo ongoing updates and validations to maintain the reference data current.

[0124] In an embodiment, a specialized Al system undergoes a rigorous initial training phase at the hands of the trained professional (or an agent thereof). This phase entails passing the articulated or written directions and concepts from the professional individual(s), such as an attorney having experience in the field of data transfers and market data contracts, to theinput of the system via an input keypad, microphone, the forwarding of data files, or other means. The body of transferred knowledge may include reference to different document types. In other more sophisticated embodiments, the knowledge may include files that can serve as documentary templates for the user. The transferred knowledge may be directed to a plurality of prospective transactions relevant to the data in question. Using this training data, the specialized Al system may employ source code to create a comprehensive and reliable model accessible to the specialized Al system for allowing the generation of accurate responses to prompts.

[0125] The training phase may in some instances be longer and more time intensive from a manpower standpoint than the traditional automated data dumps of numerous training examples for generating and minimizing complex cost functions that accompany large-scale LLM training. As noted, the specialized Al system instead undergoes a plurality of training sessions via the human interface, along with periodic updates or revisions to the domain knowledge during system operation. The system is operational to accept queries following this initial and potentially rigorous set of sessions with the legal professional. The extra time and manpower used to curate the knowledge, however, translates to responses by the system to user queries of superior reliability to its LLM counterparts.

[0126] In an embodiment, the reference data assembled during the training sessions comprises data associated with the user, domain-specific curated data, or a combination thereof. Data associated with the user can include documents generated, collected, or maintained by the user in connection with the user's specific practice or business context, such as agreements to which the user is a party, proprietary records, and internal policies reflecting the user's particular domain-specific interests; domain-specific curated data can include authoritative reference materials for the relevant domain assembled by domain experts independent of any particular user's document library, such as regulatory texts, published standards, and domain-recognized instruments applicable to the subject matter of the training sessions.

[0127] In an embodiment, the reference data assembled during the training sessions can comprise between about 50 and about 100,000 structured domain documents, for example at least 500 documents and at least 10,000 documents.

[0128] In an embodiment, the relationship between the reference data and the user is defined by the origin of the data in the user's domain context, including, for example, documents generated, accumulated, or maintained in connection with the user's business, practice, or organizational operations, such that the individual or individuals who curate orcontribute the reference data during the training sessions need not be the same individual or individuals who subsequently submit queries against the domain knowledge base.

[0129] In an embodiment, the domain knowledge base generated during the training sessions is stored in a dedicated partition associated with the user. The dedicated partition is inaccessible to other users or Al systems associated with other users, such that the knowledge curated for one user cannot be accessed by or incorporated into the knowledge base of any other user. This isolation property applies to the domain knowledge base artifact from the point of its creation during the training phase, not only during subsequent query processing. In an embodiment, the per-user dedicated partition can support confidentiality of domain-specific knowledge and the underlying document library in environments subject to cybersecurity, data privacy, or professional confidentiality requirements.

[0130] In an embodiment, the training phase includes establishing one or more scoring parameters or relevance thresholds for evaluating domain knowledge relied upon during subsequent query processing. During the curation process, domain experts or system developers configure the parameters used by the Scoring module to evaluate whether retrieved passages and documents are sufficiently responsive to a query. These relevance thresholds define the minimum standard a retrieved result must meet before the system returns a scored response. When domain knowledge processed at inference time does not yield a result meeting the configured threshold, the system returns a defined output state indicating that the domain knowledge base does not contain information sufficiently responsive to the query, rather than a scored but insufficiently supported result. In an embodiment, this output state may be presented to the user as a 'No Available Response' indicator or similar unsure-response designation. The scoring parameters and relevance thresholds established during training are applied at inference time and are not generated probabilistically per query; they reflect configuration decisions made during the curation process and can be updated in subsequent training sessions as the domain knowledge base is revised or extended.

[0131] In an embodiment, the relevance threshold for the Scoring module can be set to a value in a range from about 0.0 to about 1.0 on a normalized scoring scale, with a minimum configured threshold of at least about 0.5, such that retrieved passages and documents assigned a score below the configured threshold are excluded from the response and the system returns the no-available-response output state rather than a scored but insufficiently supported result. It is understood the range can be any multiple of 10 or a percentage. For example, 0 to 10 or 0 to 100, or 0% to 100%.

[0132] In an embodiment, the curation process produces a document-type-specific configuration template that governs how the processing pipeline interprets and analyzes documents of a given type. The configuration template encodes the interpretive rules, terminology mappings, cross-reference mappings, clause hierarchy conventions, document priority hierarchies, and processing priorities specific to a particular category of structured domain documents. In an embodiment, the configuration template captures disambiguation rules for terms whose meaning varies by document type; for example, the term "derivative" carries a distinct meaning in data licensing and copyright law contexts, where it can refer to a work derived from licensed source material, compared to its meaning in financial products contexts, where it can refer to a financial derivative instrument. By embedding this domainspecific interpretation during curation, the configuration template can ensure that the pipeline applies the document-type-appropriate meaning at each processing stage without relying on probabilistic contextual inference at query time. In an embodiment, the configuration template specifies which processing modules of the pipeline are applied, and in what sequence, when the Al system processes queries against documents of the associated type. The template is established by the domain expert during the training sessions and is applied by the Al system at query processing time.

[0133] In an embodiment, the Al system is configured to receive a sequence of successive queries concerning different issues from the user to help the user populate the relevant cited documents and schedules needed for use in a given transaction. In an embodiment, the system can subsequently assist the user in amending the documents, e.g., to incorporate desired third party input. The remaining queries and objects are managed using the specialized, pre-trained application software.

[0134] A data exchange, referred to herein as an “exchange”, is a platform or system that facilitates the sharing, trading, or transferring of data between different organizations, systems, or individuals. It serves as a centralized marketplace where data providers can publish their datasets and data consumers can discover, access, and use those datasets. The exchange's data is further valuable for price discovery in derivatives markets and as benchmarks for various financial instruments.

[0135] The Al system herein may include, among other financial data, a body of domain knowledge from a national stock exchange and a national commodity exchange for accessing, querying or otherwise using the exchange reference for different purposes, including providing automated responses to assist the entity in understanding, and entering into, a data market contract with an exchange or other financial institution. These are just two examples of many.

[0136] The legal issues and relevant document types may depend on the nature of the transaction and the posture of the financial institution and the third parties interested in accessing the data. The queries are based on a specialized Al system that uses a model including domain expert knowledge relevant to, and focused on, a domain of one or more related queries relating to the data contracts and accompanying issues. The domain knowledge, originally elicited during a training session using one or more professional experts (e.g., an attorney with longstanding and intimate experience with market data contracts and associated IP and contractual rights and obligations) may include documents or documentary templates needed for the prospective transaction, proven key concepts, and defined relationships between document types or concepts. The system may provide the documentary sources relied on for all three types of domain knowledge, as well as a display of the reasoning distilled by the knowledge graph (e.g., highlighted or underscored) text or illustrations within the source that provide specific support to the response to each query.

[0137] Curating Knowledge of the Relevant Domain. Unlike conventional chatbots / AI agents, the specialized Al system described herein may be configured to explain to a user what actions or steps the Al system is performing at a given instance in time. The specialized Al system is pre-trained (or augmented with additional training or institution specialization on the fly) to follow a series of steps to prepare a comprehensive response to a query, such as for facilitating a transaction. Because the Al system is uniquely prepared by a seasoned legal professional, the Al system can be said In an embodiment to mimic the actions that a hypothetical junior attorney may perform for the experienced attorney. These actions may also In an embodiment include preparing documents and templates, conducting legal research, and conveying an understanding of the key concepts in the query. One aspect or feature of these actions as performed by the specialized Al system is that it enables the user to understand and interpret the individual steps the agent is undertaking at a given time in the process of facilitating a transaction. In particular, understandability and interpretability represents the other half of modifiability. That is to say, if the user does not understand what actions the Al system are performing at any point in time, the user cannot then improve the system and increase its accuracy or flexibility. To that end, In an embodiment, the specialized Al agents may include a cogent explanation of the steps being undertaken when responding to user input so that an engineer or IT technician can perform any appropriate modifications to increase the system’s quality. In an embodiment, the indication of processing steps can take one or more forms, including a status message presented in the user interface, a progress display updated aseach step is completed, a log entry made accessible to the user during or after processing, a written explanation of the steps being performed, or a combination thereof.

[0138] In addition, the specialized system is capable of addressing routine initial questions, such as what meaning the system ascribes to certain terms. Because every exchange or other entity may have its own ontology, different meanings may be ascribed by the system to the same or similar terms. This definitional question is common for users initially unfamiliar with the details of a data exchange.

[0139] For this reason, a user may first query the definition of unfamiliar terms, or terms that may deviate from their ordinary meaning. Once the user poses the query, the system may curate (e.g., select, organize and oversee) the knowledge of the particular domain within which the user ran the query. Accordingly, in another aspect of the disclosure, the specialized Al system responds by curating knowledge of this market data domain so that the resulting Al systems developed and trained using the reference data are far more reliable.

[0140] In the response to a query concerning the meaning of the terminology “delayed data”, the user can interact with the UI to receive the response (that is, the definition of “delayed data” and its distinction from real-time data, for example). In addition to the response, the specialized Al system enables the user to view the supporting references (e.g., at the bottom of the screen as in FIG. 7). These references represent the sources from which the application obtained the data for use in providing the response to the query. In an embodiment, the specialized Al system advantageously enables the user to click through the references to review the exact words used by the system in preparing the response. This capability represents an innovative level of transparency that the inventors submit is needed to build a viable solution in the context of these and other high stakes domains.

[0141] As discussed further below, the list of references used by the system in curating the knowledge for the user’s query includes the name of the document, the identity of the repository, and the page(s) of the document used by the system in the response. In addition, the system makes a document viewer available on a region of the UI, such as to the right of the query response region. The user can use the document viewer and corresponding highlighted excerpts to scroll through the relevant pages from the files in the list of sources. In this manner, the user can skip the non-responsive portions of the document and can pinpoint the important sources.

[0142] The creation and use of the Al-based system is discussed herein with reference to the concepts of data usage and the scope of rights to manipulate data in the exchange. It should be understood, however, that the techniques introduced herein can be applied with equal resultsto other knowledge domains. Referring briefly to FIG. 5, for example, a user may select from a variety of domains.Knowledge Graphs. As noted, unlike conventional Al systems that make computations that rely on statistical determinations to match words and produce responsive information, the specialized Al system provides a superior approach by instead curating the knowledge of the pertinent domain. For instance, where use of the system involves the prospective legal rights and limitations of an entity to use, reproduce, sell, or otherwise exploit a given set of data or a product that relies on the data, the knowledge-curating process includes working with data provided by relevant domain experts (e.g., an attorney having a longstanding background consistent with these subjects). One objective of this process is for the legal professional and coders to create structured maps of knowledge for the domain. These maps of knowledge may include, by way of example with respect to one identified entity: identifying the key terms and concepts used in the industry at issue; determining how, if at all, the documents are related to these key terms and concepts; and concluding which documents (or excerpts thereof) the user should read for which types of topics relevant to the user.Once these activities are performed and the Al developer acquires this knowledge of curating the domain, the developer can use this knowledge to build or develop Al systems that are more efficient, reliable, and that operate to provide the user with relevant information concerning the domain of interest. In one aspect of the disclosure, the foundation of this domain knowledge base provides the Al-enhanced system with a training model that is purely deterministic and thoroughly explainable from a logical standpoint, thereby eliminating some or all of the probability-based operations currently conducted in advanced large language models like Transformers and other mainstream machine learning (ML) systems.

[0143] In another aspect of the disclosure, the domain knowledge may be maintained as a hierarchical aggregation of related data objects referenced by the inventors as a knowledge graph. While the knowledge graph is generally created and maintained based on human experts in the domain of the knowledge at issue, such as data exploitation agreements or market data contracts, it is not so limited and other legal uses are possible. The knowledge graph, or a subset thereof, is displayable on a user interface (UI) and offers an intuitive view of different features. Regions of the knowledge graph may also be selected to enable the user to elicit more data relevant to the item or object selected. The knowledge graph may also be modified, augmented, or personalized by a user and may include the features below. It should be understood that, while knowledge graphs are a powerful and flexible tool, other methods of knowledge curation are also possible and remain within the spirit and scope of the present disclosure.

[0144] Knowledge graphs beneficially provide a viewable representation of both the key data elements used in intensive data analyses systems like the Al systems introduced herein, and the pertinent relationships and hierarchies between those elements. In an embodiment, the knowledge graphs are not necessarily limited to visual representations. Rather, these tools can in alternative embodiments rely on text or voice directives. Whatever the format, the elements of a knowledge graph may be stored in a non-volatile drive such as the data repository for applications centric to a cloud. Copies can be downloaded to the workstation 102 and stored in the solid state drive, or any other form of memory, including dynamic or static random access memory (DRAM or SRAM), to enable the user to locally manipulate the elements that make up the graph. Knowledge graphs can be implemented using code executing on one or more processors (e.g., processing resources 124), or using hardware implementations. In an embodiment, the information representing the knowledge graph may be stored in a dedicated Al-enhanced device rather than a generic computing device. The user can write or revise code to further develop the graphs for the benefit on the user’s business, and the executed code can be shown in visual form on the UI.

[0145] In short, curating the knowledge of the domain includes working with the legal expert(s) to enter data that may be used by the specialized Al system to formulate relevant knowledge graphs responsive to user query. The knowledge graph can indicate visible criteria like (1) maps of the key terms used in this industry, (2) maps of how the identified documents are related to the relevant key concepts, (3) which document the user should read to analyze and assess a particular topic of concern, and (4) passages and excerpts from existing agreements. Once equipped with that knowledge, the Al system incorporates training procedures that enable the developer to transform the Al system into a deterministic one and that the response can be explained with objective logic at each step from when the query was submitted.

[0146] FIG. 3 is an example knowledge graph displayed on the user interface (UI) of a computing device. The knowledge graph of FIG. 3 shows how the various nodes operate to assist in providing a visual representation of different document types. For simplicity and to avoid unduly obscuring the concepts of the disclosure, only a subset of the number of nodes are displayed. A more practical view of the knowledge graph may be found on page 3 of Appendix 2. The key concepts associated with the knowledge graph include the document type (colored in blue), the key concepts as shaded in red, and provisions of the agreement being in yellow. The relationships between respective documents or key concepts is identified by a line between the respective nodes whose relationship is specified, along with a semi-rectangularicon. The query entered by the user in the keypad or microphone for purposes of this example is: “What is the definition of derived data?” The application may be running on one or more processors in the cloud in some examples. As in FIG. 1, the computing resources running the processors are coupled via a network connection (or, In an embodiment, locally) that enable the user to access the LLM for to perform peripheral tasks. The agreements identified in yellow include the relationships between documents and concepts identified by the text inscribed in the semi-rectangular icons. The document node “DDLA” represents the Derived Data License Agreement, and other document types related to the DDLA may include the ILA (“Information License Agreement”), Schedules 1, 4, and so on. FIG. 3 also includes various key concepts shown in violet, such as derived data, derivative works, definitions, index creation, termination, and others. The lines also provide useful information by showing that a particular section originates from the DDLA. The knowledge graph is visually superior in that the relationship icon also shows a hierarchy of documents and logical concepts. In this example, the relationship icons may identify that the subject link between a contract supersedes a schedule, which in turn supersedes a policy document. Thus, the data, documentary sources, agreement sections, statutory provisions or regulations, and key concepts are visually integrated in a manner that enables the user to understand what may constitute otherwise profoundly obscure concepts. This is in contrast to the conventional approach of using the RAG system or a conventional LLM to do a statistically-driven match and essentially “hope” that the right documents and sections are identified. It is also noteworthy that the knowledge graphs are running on a sophisticated Al platform such as a cloud platform that provides superior search capabilities, rather than merely representing a conventional keyword only search, for example.

[0147] Once the user of the system enters a query, the processing resources execute code that includes a user prompt to enter a custom query or to select from a set of standard queries generated for the particular domain selected. As an example of the operation of the system, it is assumed that FIG. 3 is being generated as a sub-graph of the domain knowledge in the domain selected by the user. Upon receiving a prompt on the UI, the user may enter a query. The query may be, in this example, “what is the definition of derived data?” The processing resources (or workstation In an embodiment) generates the graph of FIG. 3 by first identifying the key terms relevant to the query. As described further below, the dataset has been pre-trained to grasp from the query the correct ontology of the space, whether other equivalent terms may be used in the art to mean the same things, how these terms are given meaning in different exchanges or other organizations, and still other factors. Examples of key concepts may include non-display use, derivative works index creation, and derived data. For a clearer understandingof the elements of the knowledge graph and its division into constituent colors for ease of comprehension, reference is again made to the colorized knowledge graph at page 3 of Appendix 2 hereto.

[0148] After initial key terms are identified based on key concepts that are relevant to the bounds of the query, the specialized Al system may start determining the relationships between each term based on the knowledge graph based on the Al model. In the example of FIG. 3, derivative works may be deemed to be related to (302) derived data and vice versa. Index creation may also be related to (302) derived data and vice versa. Other relationships may include “read_for” (304), “Supersedes” (306), and “Section_of’ (308). Once the key concepts and relationships are established, the system continues to follow the knowledge graph to add concepts and other objects directly relevant to the query. The pre-trained model provides that “derived data” is a very related concept to “index creation”. Relevant nodes responsive to the query can be added in this manner. Once all nodes are found relating to the key concepts, the system proceeds to inquire what documents should be analyzed that relate to this question. For example, the DDL and ILA are flagged, and the relationship between the documents establish that the DDLA supersedes the ILA, which in effect provides the order of priority between the documents. The latter relationship is heretofore unknown and has been taught in the pretraining process. Unlike an LLM or RAG system, the specialized Al is system not merely guessing. Rather, it is drawing conclusions based on hard rules articulated in the training process.

[0149] Once the relevant knowledge graph of information pertaining to the query is complete, the specialized Al system then uses these identified relationships to transform the domain knowledge down to the section level hierarchy, and thereafter, the data is transformed and made available at the clause level. The clause level may be a chart that includes information regarding various definitions and their meanings, along with descriptions that provide information in direct response to the query. For the derived data query, for example, the clause level may include a summary of definitions for different available exchanges or financial bodies. These different levels result in a multi-tiered response to the query. The user can follow the generated nodes to the relevant context and select the desired object to retrieve a UI screen that includes the information at the clause level. For example, if an undesired response comes up in the graph relating to an exchange, the response can be modified and new relationships can be dynamically formed. In addition, the user can In an embodiment select a particular clause for which more information is desired. The system may provide a detailed answer (e.g.,prohibited actions regarding the use of derived data in a particular exchange) as well as clarifying terms.

[0150] While financial institutions have been the subject of examples of the present disclosure, the knowledge graph in other aspects may relate to different legal issues, tax accounting principles and other subjects. As such, the methodology as applied to market data can alternatively or additionally be applied to a wide variety of subjects above and beyond the current examples. Using the expert-based training model, in legal scenarios, the Al system is being taught to become a trained professional and to replicate the body of knowledge for the benefit of the user.

[0151] As another benefit, the specialized Al system is continuously working off a structure that can be improved, augmented, and modified over time.

[0152] In short, knowledge graphs function by internally linking words, phrases, elements, concepts and other linguistic criteria in documents, media sources, and other meanings capable of storage in memory. The linking can be performed by a developer and it can utilize any of a number of different logical constructs. Knowledge graphs are highly effective because, among other reasons, the relationships between their constituent elements with each other can be determined by one or more experts on the domain at issue. The results of the generated relationships -that is, the knowledge graph itself- can be viewed on the UI so that the structure of the relationships between data, key concepts, and relationships between documents and / or concepts, can be physically mimicked by the developer, and consequently manipulated to generate (responsive to a query) a set of maps and / or a series of meaningful logical relationships that are both highly reliable and that are germane to the current bodies of knowledge relevant to the knowledge graph and the overall Al system. Because the data underlying the knowledge graph is the ground truth, the reliability of the responses is high.

[0153] In an embodiment, the knowledge graph may utilize a hierarchical data structure. As described at length above, an instance of the data structure may be displayed on the UI using the specialized Al system. In yet another aspect of the disclosure, the user can select objects on the UI (whether by using a mouse, touchscreen, or other input means) to navigate to a level beneath the default level of hierarchy. For example, if a user clicks on or otherwise selects a particular node, the system may provide another screen of data with explanatory text detailing more specifics about the node. The explanatory text may further be accompanied by documentary sources or further explanations about the node, optionally in still different logical hierarchies. In some examples, the user can select the relationship data to obtain further explanatory information about the relationship between two nodes and which node takespriority between the two. The knowledge graph (e.g., node cluster) can be navigated by zooming in and out, panning to a specific representation of a transaction, etc. Where edits are made at any level, the modified knowledge graph can then be saved for future use by the customizing entity.

[0154] In an embodiment, the knowledge graph can be implemented as one or more of a hierarchical graph, a directed acyclic graph, a property graph, an ontology graph, a semantic network graph, a multi-relational graph, a hybrid of one or more of the foregoing, or a combination thereof. In an embodiment, the knowledge graph is a property graph, in which each node and each edge carries one or more typed labels and associated attributes that can be queried and traversed during processing. In an embodiment, the knowledge graph is an ontology graph, in which the nodes and edges encode domain-specific ontological relationships among document types, legal concepts, and regulatory provisions curated for the domain at issue. In an embodiment, the knowledge graph is a semantic network graph, in which nodes represent domain concepts and entities and edges represent semantic relationships between those concepts and entities, and the Al system can use the semantic network graph to identify conceptual proximity between a user query and relevant document provisions. In an embodiment, the knowledge graph is a multi-relational graph, in which two or more distinct edge types can exist between the same pair of nodes, allowing the representation of multiple simultaneous relationships among document provisions, statutory references, or domain concepts. In an embodiment, the domain knowledge base comprises, in addition to or as an alternative to a knowledge graph, another symbolic representation of knowledge, including a structured ontological model or rule-based logical framework, comprising a plurality of nodes and edges, or functionally equivalent structural elements, that encode the domain relationships used by the Al system during query processing.

[0155] In an embodiment, retrieving one or more portions of the domain knowledge base during traversal of the knowledge graph or other symbolic representation of knowledge is a processor-executed operation. The processor can access encoded data structures stored in memory, returning machine-readable information in the form of bytes or equivalent digital representations that encode the relevant nodes, edges, and associated attributes encountered during traversal. In an embodiment, retrieval of portions of the domain knowledge base during traversal is distinct from human reading or interpretation of document content: the processor returns stored, structured data objects from memory to the downstream processing pipeline without requiring a person to read or comprehend the underlying document text.

[0156] In an embodiment, the nodes of the knowledge graph represent one or more of the following: document types (for example, license agreements, data exploitation agreements, government contracts, securities filings, regulatory filings, or other structured domain documents); agreement provisions (for example, defined terms, grant clauses, restriction provisions, licensing limitations, or termination conditions within a contractual instrument); statutes (for example, enacted legislative texts or statutory codes applicable to the domain); statutory provisions (for example, individual sections, subsections, or defined terms within a statute); regulatory frameworks (for example, regulations, guidance documents, and oversight frameworks issued by regulatory or supervisory authorities governing the domain); compliance requirements (for example, affirmative obligations, reporting duties, or prohibitions imposed by regulatory authorities or contractual counterparties); risk factors (for example, indemnification obligations, liability exposures, breach conditions, or data use restrictions identified during knowledge curation); jurisdictions (for example, geographic regions, legal jurisdictions, national or sub-national regulatory regimes, or governmental authorities that govern the application of documents in the knowledge base); entities (for example, corporate counterparties, licensors, licensees, data providers, regulatory bodies, government agencies, or other organizational actors referenced in the document library); and concepts (for example, domain-specific legal, regulatory, or technical terms that carry defined meaning within the curated knowledge base). The foregoing node type categories are non-exhaustive, and the knowledge graph can include additional or alternative node types configured for the specific domain for which the domain knowledge base is curated.

[0157] In an embodiment, the edges of the knowledge graph can be classified into one or more typed relationship categories, each encoding a distinct type of association between connected nodes. In an embodiment, a supersession relationship indicates that one document, provision, or clause displaces or takes precedence over another, for example, a later-issued regulatory amendment taking precedence over an earlier version of the same provision. In an embodiment, a section relationship indicates that one node constitutes a section, clause, or subdivision of a parent node, encoding the hierarchical containment structure within a document. In an embodiment, a read-for relationship indicates that one document or provision provides interpretive guidance for another, such as a definitional section read in connection with an operative clause of the same agreement. In an embodiment, a related-to relationship indicates a general conceptual association between two nodes that does not fall within a more specific typed category, such as the association between a derivative works node and a derived data node as illustrated in FIG. 3. In an embodiment, a jurisdictional relationship indicates thata jurisdiction node governs, applies to, or constrains a document, provision, or entity node, encoding the geographic or regulatory applicability of a legal rule or requirement. In an embodiment, a compliance relationship indicates that a provision or document node satisfies, triggers, or maps to a compliance requirement node, encoding the connection between a contractual or operational obligation and the regulatory requirement it fulfills or implicates. In an embodiment, a risk relationship indicates that a provision or document node is associated with a risk factor node, encoding the potential legal, regulatory, or financial exposure attributable to that provision. In an embodiment, a cross-reference relationship indicates that one document or provision explicitly references another, encoding the documentary crossreference structure within or across the domain documents in the document library. In an embodiment, a level relationship indicates a connection between nodes at adjacent traversal levels of the knowledge graph hierarchy, encoding the transition from an organizational level to a section level, from a section level to a clause level, or from a clause level to a context level during traversal. The foregoing relationship types are non-exhaustive; additional relationship types can be defined for specific domains, document types, or use cases as the knowledge graph is curated and extended.

[0158] In an embodiment, the knowledge graph can comprise at least 1,000 nodes and at least 2,000 edges for a single-domain deployment, with the node count and edge count scalable, for example to between about 10,000 and about 500,000 nodes and between about 20,000 and about 2,500,000 edges, as additional structured domain documents, document types, and relationship mappings are added to the domain knowledge base during training sessions or knowledge graph modification operations.

[0159] In an embodiment, the organizational level is the highest tier in the traversal hierarchy of the knowledge graph or other symbolic representation of knowledge, encompassing groupings of documents, provisions, or entities organized at the level of corporate counterparties, legal jurisdictions, regulatory regimes, government bodies, or other organizational groupings relevant to the domain, including without limitation financial exchanges, data providers, and regulatory authorities, and traversal can begin at the organizational level and proceed downward through the section level, the clause level, and the context level. In an embodiment, the organizational level as a traversal tier is distinct from the entities node type of the knowledge graph or other symbolic representation of knowledge: the entities node type represents individual organizational actors as discrete nodes, including specific corporate counterparties, data providers, regulatory bodies, or government agencies, whereas the organizational level is a grouping-level traversal tier that can aggregate one ormore such entities and their associated documents as the top-level entry point for hierarchical traversal of the domain knowledge base.

[0160] In an embodiment, the section level is the second tier in the traversal hierarchy of the knowledge graph or other symbolic representation of knowledge, below the organizational level and above the clause level, corresponding to an individual document within the domain knowledge base or a discrete structural subdivision of a document, including without limitation numbered sections, articles, schedules, exhibits, or amendments. At the section level, traversal can identify and return the individual documents or discrete document subdivisions within the relevant organizational grouping that contain information responsive to the query.

[0161] In an embodiment, the clause level is the third tier in the traversal hierarchy of the knowledge graph or other symbolic representation of knowledge, below the section level and above the context level, corresponding to individual provisions, defined terms, operative statements, obligations, or restrictions within a section. At the clause level, traversal can identify and return the specific clauses or definitions within a relevant section that are directly responsive to the query, including clause-level content such as summaries of definitions, permitted uses, prohibited uses, or other discrete units of meaning within the section.

[0162] In an embodiment, the context level is the lowest tier in the traversal hierarchy of the knowledge graph or other symbolic representation of knowledge, below the clause level, corresponding to the specific textual passages, excerpts, or semantic units within a clause that are directly responsive to the query. At the context level, the Al system or hybrid Al system can retrieve the granular textual content underlying a clause-level response and surface that content to the user with source identification, enabling per-passage traceability to the domain knowledge base.

[0163] The documents may vary widely depending on the intended application and may include contractual instruments like license agreements as well as related agreements, or schedules and exhibits that may accompany one or more of the agreements.

[0164] Unlike conventional search engines or sophisticated Al systems in the marketplace, the disclosed Al system is configured to provide responses and information that are not only specialized to meet the needs of a system user, but that also provides a high degree of reliability in the fields of data exploitation and related legal concepts not heretofore achieved by conventional machine learning solutions.

[0165] In the course of navigating the knowledge graph, the user can also select icons to determine both the exact documentary sources relied on to populate the icon with a label andinformation, but also highlighted excerpts from those sources to direct the user right to the pertinent information.

[0166] As noted, even today’s impressive AIs are inadequate for providing responses to such questions, as the questions are too fact and context specific to enable an entity to reliably use responses from a general purpose LLM. Much of what a user sees with existing applications using LLMs include information useful for navigating to a demonstration or referral to a related application that may be textually streamlined and well -articulated, but that unfortunately is not very accurate. While LLMs continue to improve, even state-of-the-art LLM may provide the user with correct answers 85% or 90% of the time. Otherwise, these LLMs may hallucinate such as by confidently producing false material, they may present inaccurate data, and they may fabricate “facts”, meaning that for delicate or instrumental legal questions that are involved in data sciences and related legal instruments, LLMs are insufficient. The legitimate exploitation of data and related matters in the data science domain more than likely involves very high stakes, including institutions entering into multi-million dollar decisions based on these contracts.

[0167] Unlike general Al agents or chatbots that may rely on LLMs trained across massive bodies of knowledge, the Al system herein is based on pre-trained data from one or more experienced professional individuals, such as a lawyer with a long and intensive history or mediating transactions in the field of data exploitation.

[0168] As described above, knowledge graphs function by internally linking words, phrases, elements, concepts and other linguistic criteria from documents, key sections of those documents, media sources, and other meanings capable of storage in memory. The linking can be performed by a developer and it can utilize any of a number of different logical constructs. Knowledge graphs are highly effective because, among other reasons, the relationships between their constituent elements with each other can be determined by one or more experts on the domain at issue and the results of the generated relationships -that is, the knowledge graph itself- can be viewed on the UI so that the structure of the relationships between data, key concepts, and relationships between documents and / or concepts, can be physically mimicked by the developer, and consequently manipulated to generate responsive to a query a set of maps and / or a series of meaningful logical relationships that are both highly reliable and that are germane to the specific bodies of knowledge relevant to the knowledge graph and the overall Al system.Query Processing Pipeline

[0169] In another aspect of the disclosure, the Al system is also coupled to an LLM. For example, the computing device (laptop, workstation, etc.) or, alternatively, the processing resources in the cloud (and accessible to the computing device), are configured to perform collateral functions that may be pertinent to the user’s queries but that do not substitute for the pre-trained data originating from the one or more legal experts.

[0170] In an embodiment, the LLM performs only peripheral tasks that are ancillary to the core processing functions of the Al system. These ancillary tasks can include clarifying the intent of one or more queries submitted by the user prior to processing against the domain knowledge base, and confirming the relevance of returned objects or passages following such processing. The LLM does not perform core document interpretation, traversal of the domain knowledge base, or reasoning against the curated domain documents; those functions are performed by the deterministic, knowledge-graph-based processing architecture.

[0171] FIG. 4 is a conceptual diagram 400 that provides a logistical overview of how the domain knowledge feeds into the specialized Al system to answer complex legal questions in accordance with an aspect of the disclosure. The process begins when a user enters a query using the keypad or other input instrument. In the case where the application is maintained in the cloud and its processing resources, the query is uploaded to the cloud. In an embodiment, the application software may reside on a workstation, personal computer, or laptop of the user. Leveraging the pre-trained data, at module 404, the system may refine the user’s query to more closely align with domain-specific terminology, particularly in the case where the body of knowledge maintains its own ontology. In an embodiment, the Question Enrichment module 404 can perform this alignment as an active transformation step, producing one or more enriched queries as intermediate artifacts. Each enriched query incorporates the domainspecific language of the domain knowledge base in place of the original query terminology, and the enriched queries, not the original user query, serve as the inputs to the downstream processing modules. This transformation is distinct from the background terminology mapping maintained by the Handle Ontology / Definitions module 402: the Ontology / Definitions module 402 establishes and maintains the domain's terminology mappings, while the Question Enrichment module 404 applies those mappings to actively rewrite the user's query into a domain-aligned form before document filtering and graph traversal begin. Optionally, the user may access the dialogue box for the LLM to ensure that the summary of the user’s query as entered into the specialized Al system is accurate. In this case, the user should be cognizant about the above-referenced limitations on LLM systems and should ensure that the ultimatequery comports with the user’s specific request. Upon receiving the query, in the documentfiltering module 406, the processing resources of the Al system may select relevant documents based on the curated metadata and on the definitions that relate to the subject of the query.

[0172] The Question Enrichment module 404 achieves terminology alignment through a deterministic, rules-based process. The module identifies terms in the user's original query that differ from the vocabulary established in the domain knowledge base, including terms that carry domain-specific meanings distinct from their general usage. For each such term, the module applies the terminology mapping derived from the domain ontology maintained by the Handle Ontology / Definitions module 402, substituting the corresponding domain-specific term or phrase for the original query term. In an embodiment, the Question Enrichment module 404 can also perform query expansion, replacing abbreviated or colloquial references with their full domain equivalents as established during the curation phase. In an embodiment, where a query term is associated with multiple possible domain-specific meanings, the module can apply disambiguation rules curated during the expert training phase to select the interpretation most consistent with the overall query context. Each transformation step is governed by rules and definitions established during curation of the domain knowledge base, consistent with the deterministic processing architecture of the Al system.

[0173] As shown by the Handle Ontology / Definitions module 402, the question enrichment module 404 and document filtering modules 406 are also responsible for ensuring that the correct ontology of the data and the applicable definitions are being formed. In an embodiment, following document selection by the document filtering module 406, the traversal module traverses the domain knowledge base from an organizational level to a section level to a clause level to a context level to identify and retrieve the relevant portions of the domain knowledge base responsive to the enriched query, as described herein. In an embodiment, the traversal module can traverse one or more levels of the domain knowledge base hierarchy, including any combination of an organizational level, a section level, a clause level, and a context level, to identify and retrieve relevant portions. The retrieved portions are provided as input to the Chunk Extraction module 408. Using the selected documents and the domain knowledge, the system may thereupon extract relevant sections of text, including without limitation definitions, regulations that govern parts of the intended transaction, and clauses from the agreement(s) at issue. At Chunk Extraction module 408, the system (here the processing resources) then segments those extracted clauses and relevant sections of text for a more detailed analysis of the text.

[0174] With continued reference to FIG. 4, the refined chunks and any analysis results are provided to the Ordering and Document Relationships module 412. In the fields of data exploitation and market data agreements, the identified agreement may include exhibits and schedules. Moreover, one or more key concepts may be related to the subject agreement(s). The specialized Ordering and Document Relationships module 412 of the Al system identifies necessary cross-references (Handle Cross Reference module 410) between documents and / or concepts and resolves clause hierarchies (Handle Order of Interpretation module 414). For the latter operation, one or more clauses in an agreement supersedes other clauses, and this module is responsible for resolving these hierarchies such that the user can know which clause prevails in the event of a conflict or ambiguity. In the example case shown, the ordering module (and possibly the chunk extraction module) ensures that the cross-references relevant to the prospective transaction are managed such that cross-referencing information is made available as part of the domain knowledge. Referring still to the Ordering and Document Relationships module 412, the specialized system uses the expert dataset model to identify and delineate the order of interpretation of the agreement provisions (or in some cases, other data including regulations and statutes). The output from this module is passed to the Synthesis Module 416. Here, the specialized Al system may compile the domain information into a coherent response. The processor(s) executing the relevant code may reference the domain knowledge to ensure that the reasoning of the question is sound at the Handle Reasoning module 418. (In an embodiment, after reviewing the response, the user may resort to the LLM where necessary to re-summarize the information. In this example, however, care must be taken to ensure, if an LLM is consulted, that the reasoning and logic used by the pretrained model is not superseded, inadvertently or otherwise, by the output information from the LLM. This ensures that the synthesized data retains the accuracy inherent in the ground truth from the pre-trained model of the specialized Al system). Thereafter, the Scoring module 420 evaluates each response for accuracy and confidence, ensuring only reliable outputs are returned in the response. See FIG.7.

[0175] The process flow introduced in FIG. 4 and relevant to the user query is not limited to any specific implementations, and other transactional methods involving different data may be used. For the purposes of continuity, this disclosure will now focus on the example use case of market data contracts, in which a legal expert has trained the system (and provided any subsequent material updates to maintain the domain knowledge current).

[0176] Based on this processing chain, the Al system can generate in a Knowledge Artifact module 422 a knowledge artifact created by human experts in the relevant domain, such as adocument with reader-friendly summaries, text, any tables, and other objects that may be responsive to the user’s query. As described below, the documentary sources upon which the synthesized data relies are also provided to the user in this example, as well as specific citations to the key passages in those sources. Advantageously, because the substance of the knowledge artifact is created by human experts in the relevant knowledge domain, rather than passages that result from statistical algorithms used by LLMs, reliability of the response is significantly increased.

[0177] In an embodiment, the pipeline processes queries against documents of a given type according to the document-type-specific configuration template established for that document type during the training sessions. The configuration template determines which processing modules are applied, in what sequence, and according to what interpretive rules when the pipeline encounters documents of the associated type.

[0178] In an embodiment, the processing pipeline of FIG. 4 applies, in sequence, (1) the ontology and definitions module, (2) the question enrichment module, (3) the document filtering module, (4) the traversal module, (5) the chunk extraction module, (6) the ordering and document relationships module, (7) the synthesis module, (8) the reasoning module, and (9) the scoring module, while other embodiments may employ any one or more of these modules in any combination.

[0179] In an embodiment, the Al system can generate a plain-language response to a single query processed through the pipeline within about 1 second to about 300 seconds, depending on the size of the document library and the complexity of the query, with matrix processing operations generating responses for at least about 10 document-query cell combinations per second for document libraries comprising up to at least about 10,000 documents. It is understood the speed of the response is directly related to the number of processors and memory available to carry out the process.

[0180] Having addressed the basic process flow, an exemplary implementation is now considered involving market data ownership contracts (or market data contracts”) using a traditional RAG. In the exemplary scenario depicted, an organization is tasked with reviewing multiple market data contracts sourced from various exchanges. Initially, the conventional technique of using a RAG is first explored, followed by the contrasting highly selective knowledge embedded in the system of the present disclosure. The user of a traditional RAG may retrieve clauses based on keyword matching. Examples may include the RAG’s retrieving of concepts like “ownership” and “IP rights” based on the user’s initial input, while altogether missing context that may be found in cross-references of documents or definition sections ofan agreement. As a result of this incomplete and imprecise analysis, the RAG may proceed to generate incomplete or erroneous summaries as responses. As noted, these summaries fail to capture items like subtle difference in the definition of terms and how the contractual obligations may be erroneously represented as a consequence.

[0181] Referring instead to the use of knowledge-domain systems of the present disclosure, these systems can capably identify and extract precise definitions from each agreement at issue. For market data contracts, the specialized Al system can, for example, (i) identify and extract the precise definitions of terms like “data”, “derived information” and “proprietary data” from each agreement, (ii) map out accurate cross-references including references to earlier clauses, notice period requirements, and usage limitations on the data, and (iii) ensure the presence of consistent interpretations across multiple sources while reconciling any differences in ontologies. Examples of the latter include, as noted, an output reconciling different definitions, including whether and to what extent the definitions are consistent with each other. When various definitions diverge or are not present in one of the agreements, the specialized Al system can point out these issues and their ramifications on any transaction. The result is a prompt output including accurate capturing of subtleties and a legitimate representation of pivotal definitions and requirements.

[0182] As is evident from this example, knowledge curation according to the present disclosure offers a framework for addressing the inherent limitations of traditional RAG systems, especially where, as here, high-stakes legal domains requiring precision are paramount. By systematically capturing and validating the nuances and cumbersome information in most legal documents, the aspects disclosed herein ensure that the Al system delivers accurate, reliable, and context-aware outputs.

[0183] Alternative embodiments arise from the use of knowledge curation. In yet another aspect of the disclosure, scaling curation processes can be introduced to accommodate the size and needs of an organization. Given that the corpus of legal documents is undergoing continuous evolution and systematic expansion, the specialized Al system in an embodiment includes an engine that automates aspects of expert validation and annotation. Such automation can include automated prompts for subsequent training by the legal expert and can encompass validation of the new training input across different data repositories to ensure its integrity.Query-Response Interaction and Source Attribution

[0184] FIG. 5 is an example screenshot of a user interface of the specialized Al system according to an embodiment. The system at its core is a specialized query / response system. The left side of the screen provides a dialog box for search entries as well as a view of a pluralityof different exchanges. Each data entry includes a data exchange or institution with files provided by the legal professional in the training phase. In the example shown, the CME menu was expanded. As shown by the expanded menu and bold font, the CME_device-fee- waiverpolicy. pdf document was selected by the user. As a result, documents corresponding to device fee waivers and Schedule la governing the CME Distribution of Information is portrayed on the right side menu. The “Library” icon on the left has been selected for this view. The user can navigate through the documentary sources and can search the files, or develop a custom query.

[0185] FIG. 6 is an example display screen providing a plurality of CME-specific queries from which to choose and a Q&A section with a response based on the licensing requirements in an Alternative Trading System (ATS), in accordance with an embodiment. This screen enables the user to select a query from a list of common queries, as displayed on the left side. It is noteworthy that the icon “Matters” is selected for this view. It is assumed that the second query relates to further documentary sources for the CME Documents selected from FIG. 5. After the query related to “licensing requirements for using CME marketing data in an ATS” is selected from the list of queries, the Q&A section on the right side of the screen provides a response with descriptive information squarely relevant to the query. The user then can “Load detail”, for example, to identify the documentary sources for the descriptive response.

[0186] In an embodiment, the system may present the user with a domain-specific query library in addition to a free-form query input. In an embodiment, the domain-specific query library is a pre-built collection of standard queries curated in connection with the domain knowledge base and configured to address the principal subj ect matter and analytical obj ectives of the domain. The query library can be assembled during the knowledge curation process and updated as the domain knowledge base is refined. FIG. 6 illustrates an example of such a domain-specific query library, in which a list of selectable pre-configured queries associated with the CME market data domain is presented on the left side of the screen. In an embodiment, the user may select one or more queries from the domain-specific query library, enter one or more free-form queries, or use a combination of both input modes.

[0187] FIG. 7 is an example display screen providing the response to the ATS query of FIG. 6, and in particular direct access to documentary sources supporting the response, with the key text highlighted in this example via a boundary, together with the scores of different document sources. In this perspective, the response is provided on the left side of the screen. The right side displays the relevant text of documentary sources upon which the legal professional relied in providing the response. The “highlighted” paragraphs on the right sidereproduce the specific paragraphs relied upon by the expert in providing the response. For purposes of this example, the highlighted references from the first of three pages (“Page 1 de 3” on the upper right corner) include the text selected by the boundaries, which is ordinarily highlighted in yellow. In addition, the relied upon documents and their page citations are reproduced at the bottom of the screen with the scoring of relevance determined via FIG. 4, as well as the citations to the page and line relied upon.

[0188] In an embodiment, the response can include, for one or more passage relied upon in generating the response, per-passage score types indicating the relevance of that passage to the query. In an embodiment, the score types rate accuracy of the response as well as score the productivity indicating respectively the relevance of that passage to the query and the productivity of the system relative to the speed of the response. Per-passage score types can be presented as a Ul-visible element associated with each highlighted passage in the document viewer, for example as a numerical value, percentage indicator, or categorical ranking displayed adjacent to each highlighted passage excerpt. A per-passage score type is distinct from the document-level scoring of relevance: the document-level score reflects the overall relevance of a source document to the query, while a per-passage score type reflects the relevance of a specific extracted passage within that document. In an embodiment, the response display can present both document-level scores and per-passage score types together, providing multi-tier transparency into the evidentiary basis for the response and enabling the user to distinguish high-confidence source passages from lower-confidence supporting passages without reviewing the full source document.

[0189] In an embodiment, the Al system returns an unsure response when the curated domain knowledge base does not contain information responsive to a query. The unsure response is a defined output state of the system, distinct from a low-confidence response: the system generates the unsure response when no passage in the user's document library satisfies the relevance threshold established during the knowledge curation phase for the query at issue. In an embodiment, the unsure response can be presented to the user as a "No Available Response" indication in the query-response interface. In an embodiment, the unsure response state can apply at the response level for a single-query interaction and at the cell level in a matrix output; for example, in the multi-document, multi-query matrix illustrated in FIGs. BASE, a cell that finds no responsive document or passage can display a "No Available Response" indication rather than a plain-language response and citation.

[0190] In an embodiment, the response generated by the Al system can be transmitted to another user, a computing system, or a combination thereof. A computing system that receivesthe response can include an automated downstream processing system, an application connected to the Al system via an application programming interface, or a third-party computing service configured to receive and further process the response.

[0191] In addition to the available screen shots described above, the user can formulate a custom query that can be modified and refined to suit the user’s interests. The domain knowledge will form the objects and links required for the response.Domain Extensibility

[0192] Using the domain knowledge described herein as a basis for the ground truth training enables related laws to be explored. Determinations can be made to help ensure that the user is properly using the obtained data after the fact. For example, one law with significant links to data exploitation matters is the copyright law.

[0193] In the data usage domain, copyright is ubiquitous. One issue that arises in copyright law is whether an entity has the right to create derivative works using a particular document or source of data. Another issue may include whether an entity has a claim against another for copyright infringement by virtue of the alleged infringer creating derivative works of the author’s copyrighted material. For this reason, the concept of a derivative work may be of particular importance to a company looking to exploit an original work (e.g., with an expired copyright, etc.) with multiple derivative works. The same concept may be equally important to an entity who may suspect that its own documents or works of authorship have been infringed by the fact of other entities creating derivative works. Within this example, if the user enters a query including the term “derivative work”, the knowledge graph may capture important information and documents relating to the subject, not the least of which is which an identification of documents the user should be reading to answer questions pertaining to whether the derivative works at issue are a cause for concern. This sequence is another example of the benefits of curating the data as described above.

[0194] The use of a “derivative work” may also be understood and applied in other contexts, such as whether a license agreement should include or exclude derivative works, and under what circumstances the derivative works are owned by one of the parties to a license agreement. Detailed information regarding the derivative works is found in the ground truth dataset that was generated by the relevant expert(s) as further described herein.

[0195] The documents may be organized in different ways, depending on the implementation. For example, in an embodiment, the data in the knowledge graph may refer to or identify set of contracts, along with any appendices and schedules to those contracts. Contractual data may involve a trademark license agreement and any relevant exhibits. Inanother embodiment, the contractual documents may be combined with the regulatory documents, such that the user can access either type of document using the query. These illustrations are exemplary in nature, and the knowledge graph may incorporate a number of other conceptual data types without departing from the scope of the disclosure.

[0196] As noted, a tradeoff is present between the new and conventional systems. In the new system proposed herein, the developing entity takes time to work with the relevant expert and formulate the knowledge graphs. However, at the end of the initial training process, the Al system is able to work off a level of accuracy that is much higher, because the Al agent is working off the ground truth. Conventional Retrieval Augmented Generation System-based technology, which combines name searches with meaning searches, lack the precision of the new technology, at least because the cost functions are not accurate enough to capture the ground truth when the models are trained. The proposed Al system may be used to capitalize on the portions of conventional large-language models (LLMs) that perform best. For example, certain LLMs may be highly proficient at summarizing data, or answering targeted questions about large amounts of information in a way that no typical human or basic computer system can do. That said, the new Al system proposes abstaining from use of large language models where they have to make a logical conclusion that is contrary to the domain knowledge. The new Al system endeavors to work in the above areas where LLMs typically do reasonably well, while using personal knowledge and expert-initiated skills which the inventors have gained from applying their cutting edge technology in high-stakes situations (e.g., trading). Thus, in aspects of the new Al system, the use of probabilities is limited to other Al-based functions to ensure that any use of the LLM will not affect the integrity of the trained information provided by the legal professional. For applications requiring the levels of accuracy commensurate with the examples above, the new Al system can employ the requisite curate knowledge to set the ground truth and ultimately enable highly reliable, logical decisions to be made, including in the areas of the rights to exploit certain data.

[0197] The backdrop of the subset of the knowledge graph shown also applies to the example user query, above: “What is the definition of derived data?” The proposed Al system operates on the knowledge graph by considering the ontology of the relevant terminology as noted above. The algorithm proceeds to identify the key terms relating to the query. The system identifies the nodes corresponding to the derived data concept. Thereupon, the system uses the knowledge graph to identify related key concepts. One such concept common to derived data is index creation (which is part of the ground truth from which the system was trained). The system identifies all remaining nodes in the knowledge graph related to derived data. Forexample, the system of FIG. 3 may identify the DDLA document and may further indicate that the relationships between the derived data and DDLA indicate an order of priority to be followed. Once the DDLA is identified, the relationship link between the DDLA node and the ILA node demonstrates that the DDLA supersedes the ILA. Thus, in this context, the system determines that the ILA should be read and the appropriate links should be incorporated into the response to be provided to the user. It is noteworthy that the Al system described herein, including the relationship between the DDLA and the ILA, was derived comprehensively during the curation process leading to the ground truth, rather than being based a consequence of statistical probabilities as is characteristic of existing Al systems. As noted, the process shown in FIG. 4 may begin at the section level and proceed by following the generated rules to eventually analyze the agreement at the clause level. A tree of nodes can be systematically followed down to the context level. The system combines and integrates all related concepts relevant to the query using the coded rules characteristic of the knowledge graph.

[0198] In an embodiment, the knowledge graph architecture described herein is domainagnostic. The curated knowledge base approach, comprising expert-guided training, hierarchical knowledge graph traversal, deterministic query processing, and plain-language response output with identified source passages, applies to any domain in which complex, structured documents require precise interpretation. In an embodiment, the domain is financial services and market data contracts.

[0199] In an embodiment, the document library comprises regulatory filings, statutory provisions, agency guidance documents, or a combination thereof. The knowledge graph can be structured to represent the relationships among applicable regulations, the compliance obligations they impose, and the provisions within the document library that map to those obligations. The Al system, for example, can enable a user to query a regulatory document corpus to determine which provisions apply to a specific transaction or activity, whether particular documents satisfy applicable requirements, or which regulatory authority has jurisdiction over a given matter.

[0200] In an embodiment, the document library comprises cybersecurity policies, data protection regulations, vendor agreements containing security -related provisions, or incident response documentation. As illustrated in FIG. 5, a Cybersecurity Regulation domain category can be loaded into the system as a distinct knowledge domain with associated documents. The knowledge graph can be structured to represent the relationships among applicable cybersecurity frameworks, the obligations they impose, and the specific provisions within the document library that address those obligations. The Al system, for example, can enable a userto identify provisions across a portfolio of vendor agreements that present cybersecurity risk exposure, or to determine whether an existing policy satisfies the requirements of a designated cybersecurity framework.

[0201] In an embodiment, the document library comprises contracts, regulatory materials, or supporting documentation organized by counterparty jurisdiction or geographic region. Knowledge graph nodes can represent jurisdictions and geographic entities in addition to document types and agreement provisions. The Al system, for example, can process a multijurisdiction document corpus to generate output organized by jurisdictional grouping, for example a matrix whose organizational-level axis corresponds to counterparty jurisdictions, enabling a user to compare rights, obligations, and restrictions across a portfolio of agreements spanning multiple geographic regions or regulatory environments.

[0202] In an embodiment, the document library comprises statutory texts, legislative history, government contracts, securities filings, loan agreements, customer subscriber or service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, or a combination thereof. The domain knowledge base can be curated to incorporate the interpretive relationships among statutory provisions, implementing regulations, and agency guidance, enabling the Al system to respond to queries about how a statutory requirement applies to a specific transaction or activity. Government procurement agreements, securities registration statements, and loan and financing documents represent additional structured document types to which the same curated knowledge base approach applies, with domain-specific curation tailored to the terminology and interpretive conventions of each document category. The domain examples described herein are illustrative and non-exhaustive.

[0203] In an embodiment, the Al system can process queries submitted in multiple languages and can return responses in the language of the query or in a user-specified language. In an embodiment, the domain knowledge base can be curated from domain documents in multiple languages, with language-specific ontologies configured to capture the terminology, definitions, and cross-reference structures characteristic of each language. In an embodiment, the language-specific ontology configuration can enable the system to resolve domain-specific terminology across language boundaries within the same knowledge graph, supporting use cases in which the document corpus spans multiple jurisdictions, regulatory regimes, or languages.Knowledge Graph Modification

[0204] A significant benefit of the system described herein relates to its ability to improve bodies of complex data structures over time. As one example, a user working with a particular exchange may not be receptive to one or more aspects of the response returned based on the query. The user may note, for instance, that the response failed to include a document deemed relevant in the domain to which the exchange relates. At this point, the user can modify the knowledge graph by adding, deleting, or changing one or more nodes to include additional documents, specify further operations, or remove unnecessary functions, among other modifications. This benefit is particularly relevant to documents that are widely used, such as specific bodies of well-known regulations, e.g., those relating to tax-based implementations, intellectual property regulations, and the like. In short, the knowledge graph is not a stagnant data structure, but it is versatile and can be modified to fit the needs and standards of an organization. The knowledge graph can be improved and refined over time to consider the needs and preferences of different users or different data domains. In contrast to existing Al solutions, the system described herein can use the knowledge graphs originally created to make changes on the fly and to progressively increase the accuracy and utility of the system.

[0205] It is also noteworthy that, once the system can reliably provide responses to queries about a particular data source, such as a contract, users can build additional workflows into the knowledge graph. These additional workflows may be used to develop new and more sophisticated products that can be applied to different exchanges. Where the user wants to use these products such as inputting a query about the operation of an exchange such as a prohibition or restriction on the right to perform some action, the user may receive summarized text relating to the exchange. The user may then select the text to reveal the underlying details specific to the exchange at issue

[0206] In an embodiment, a summary grid is provided that includes a response to the query along with supplemental information about the query like the different possible use cases, the requested definitions (e.g., derived data as in the above example), and related information such as relevant attributes concerning the topic of the query. In an embodiment, as requested by the user, the response includes answers to the question over a plurality of different exchanges. The user can traverse the graph and identify specific responses, e.g., relevant to a specific exchange. The user can navigate to the details and sources of the response, such as the details concerning conditions relevant to a particular exchange. In the example query asking for the definition of derived data, the summary grid in a responsive UI screen can include related definitions, attributes and risks relevant to a complete response. These include the frequency of the data,the definitions of derived data and other definitions that the user may need to understand the response, permissions, prohibitions, policies concerning index creation, identically terms used differently in different domains, etc. In one configuration, a document viewer is prompted responsive to the user selecting a particular response. The document viewer may include conditions, restrictions, or assumptions relevant to that response.

[0207] The above procedures illustrate how the specialized Al system of the disclosure can respond to the user’s actions by consulting the underlying knowledge graphs representing the kernel of the system, and to create new solutions to different or additional problems using the knowledge graphs as a base. One such example may be a benchmark study conducted by an entity that may be relevant to different exchanges. Depending on the purpose of the study and the results obtained, the entity may wish to create a response that considers restrictions, conditions, or other criteria from the study and to incorporate the criteria into a new solution relevant to the applicable exchanges.

[0208] In short, the Al system described herein can be realized by the creation and use of knowledge graphs to establish the initial Al ground truth at a high level of reliability, and the prospective modification of one or more aspects of the knowledge graph during operation of the Al system to enable the system to consistently increase its accuracy and to provide new informational solutions to the user based on the specific knowledge domain.

[0209] In an embodiment, the Al system is configured to simultaneously process one or more queries against a plurality of documents in the document library, generating responses in parallel across the document corpus. This simultaneous multi -document processing capability can be applied to document corpora of varying scale, including corpora comprising at least ten documents, at least one hundred documents, at least one thousand documents, at least ten thousand documents, or at least one hundred thousand documents. In an embodiment, the Al system processes multiple distinct queries simultaneously against the same document corpus, generating a response to each query across all documents in a single processing operation.

[0210] In an embodiment, the Al system can simultaneously process at least 2 distinct queries against the document library in a single processing operation, for example between about 2 and about 100 simultaneous queries, with the resulting responses organized into a matrix comprising between about 4 and about 1,000,000 document-query cell combinations.

[0211] In an embodiment, the response to a multi -document, multi-query processing operation is provided in the form of a matrix. The matrix comprises a plurality of cells, each cell corresponding to one or more documents in the document library. One axis of the matrix, whether rows or columns, corresponds to one or more of the queries submitted by the user, andthe other axis corresponds to groupings of documents at a hierarchical level, including an organizational level, a section level, a clause level, a context level, or a combination thereof. In an embodiment, organizational-level groupings include individual corporate entities, counterparties, jurisdictions, or regulatory domains. In an embodiment, the row and column orientation of the matrix is transposable, permitting the user to invert the axis assignments and view the same underlying response data from a different organizational perspective without reprocessing.

[0212] In an embodiment, each cell of the matrix comprises a plain-language response to the applicable query with respect to the one or more documents corresponding to that cell. In an embodiment, each cell further comprises one or more passage identifications identifying the specific document, section, and passage relied upon to generate that cell's response, providing the user with per-cell traceability to the source material underlying each response in the matrix. In an embodiment, the response associated with each cell further comprises a per-passage score type for each passage relied upon, indicating the relevance of that passage to the query. Per-passage score types are a per-cell output element distinct from any aggregate document-level scoring performed by the processing pipeline.

[0213] In an embodiment, each cell of the matrix is individually inspectable. The user can select a cell to access a cell-level view comprising the plain-language response, the one or more passages relied upon, the documentary sources from which those passages are drawn, and any per-passage score types associated with those passages. In an embodiment, the cell-level view includes a document viewer providing access to the full source document with the relied-upon passages highlighted, implementing the trust-but-verify architecture at corpus scale.

[0214] In an embodiment, for any cell in the matrix for which no document in the corresponding portion of the document corpus contains information responsive to the applicable query, the Al system returns a defined output state for that cell indicating the absence of responsive information. This output state is not a low-confidence probabilistic estimate, it is a determinate output of the deterministic processing architecture, reflecting that the domain knowledge base does not contain information sufficient to answer the query with respect to the documents corresponding to that cell. In an embodiment, the output state is presented to the user as a discrete cell-level indicator, distinguishable from cells that contain a substantive plainlanguage response. This per-cell absence indicator enables the user to identify gaps in corpus coverage at the document or document-group level without the system generating an unsupported response. In an embodiment, the defined output state is labeled "No Available Response" or an equivalent indicator communicating the absence of responsive information.Dashboard and Portfolio Analytics

[0215] The Al system described herein can be configured to provide a portfolio analytics dashboard as a distinct output mode from the individual query-response interface. In an embodiment, the dashboard presents aggregate intelligence synthesized across a plurality of structured domain documents in the document library simultaneously, enabling users to assess portfolio-level liability exposure, identify non-conforming documents, and / or obtain classified analytical views of an entire document corpus. The dashboard is a creation of a new analytical artifact that transforms complex, cross-document information into actionable intelligence at the portfolio level.

[0216] In an embodiment, the dashboard is organized into a plurality of navigable analytical views accessible via a tabbed interface, as illustrated in FIG. 9A. The views can include, for example, a client liability profile snapshot, a counterparty profile snapshot, a counterparty vs. issuer / guarantor view, and / or a geographical analysis. Each view presents a distinct analytical perspective on the same underlying document corpus, and the user can navigate between views without resubmitting queries to the Al system.

[0217] In an embodiment, the visualization of the response to a multi -document, multiquery processing operation comprises a domain-specific visualization corresponding to a type of domain associated with the one or more queries. The Al system can select the visualization type based on the domain type of the query, such that a query directed to the geographical distribution of counterparty obligations across the document corpus can generate a geographical analysis visualization, a query directed to liability exposure of the document corpus can generate a liability profile visualization, and a query directed to data rights provisions within the document corpus can generate a data rights image visualization, as illustrated in FIG. 9A. The correspondence between domain type and visualization type is a configurable, non-exhaustive mapping that can be extended to additional domain types and visualization types beyond those described herein.

[0218] In an embodiment, the client liability profile snapshot view presents a portfolio summary panel displaying aggregate document counts and classification metrics across the document library, along with a configurable date filter enabling time-bounded analysis of portfolio metrics. In an embodiment, the view further presents a non-conforming contracts table identifying documents within the library that do not conform to standard terms. The table can present, for each non-conforming document, the entity name, contract classification, originating jurisdiction, contract and expiration dates, and open exposure value. In anembodiment, the user can select a specific entry in the non-conforming contracts table to access additional details associated with the selected document, as illustrated in FIG. 9G.

[0219] In an embodiment, the dashboard includes an unexpired liability exposure panel presenting an aggregate view of liability classifications across the active documents in the document library. The Al system can classify document provisions into one or more liability categories and aggregate the classified provisions across the document corpus to generate the portfolio view. The liability categories can include, without limitation, confidentiality obligations, intellectual property infringement provisions, gross negligence clauses, indemnification obligations, payment obligations, privacy provisions, and security breach provisions. In an embodiment, the panel presents the liability classification distribution as a visual chart and includes a user-selectable filter enabling display of a subset of liability categories, as illustrated in FIGs. 9B-9F.

[0220] In an embodiment, the dashboard further includes an expired liability analysis panel presenting comparative analytics on liability provisions derived from expired documents in the document library. The expired liability analysis enables users to assess historical liability exposure and identify patterns across document cohorts over time.

[0221] In an embodiment, the dashboard further includes a client insurance coverage panel presenting risk status metrics derived from the document library. In an embodiment, the risk status metrics are displayed with numerical counts and comparative indicators enabling the user to assess coverage status across the document portfolio.

[0222] In an embodiment, the Al system stores dashboard output, portfolio analytics results, and query responses on a local storage device associated with the user, on a cloudbased storage service, or a combination thereof. In an embodiment, the stored output is retained within a dedicated partition associated with the user and is inaccessible to other users, enabling retrieval and review of prior analytics sessions.Method of Use

[0223] FIG. 10 is a block flow diagram depicting an embodiment of the method. As illustrated, the method comprises receiving, from a user, one or more queries associated with a document library comprising a plurality of structured domain documents (block 210); processing the one or more queries using an Al system or hybrid Al system and a curated domain knowledge base (block 220); and providing, to the user, a plain-language response with source identification (block 230).

[0224] In an embodiment, the processing step of block 220 comprises the following sequence of operations. The Al system or hybrid Al system recognizes the document type ofone or more documents in the document library (e.g., Recognize document type) and recognizes a document sub-type of those documents (e.g., Recognize contract type). The Al system or hybrid Al system selects one or more documents from the document library responsive to the query (e.g., Select document responsive to query) and reads those documents in their entirety (e.g., Read document in entirety). The Al system or hybrid Al system identifies the existence of any addenda and attachments associated with the selected documents (e.g., Identify existence of addenda and attachments) and reads the names of the parties to those documents to determine which party is making the query (e.g., Read names of parties). The Al system or hybrid Al system then recognizes the order of priority for interpretation of and response to the query (e.g., Recognize order of priority) and reads the selected documents according to document type and document sub-type (e.g., Read document according to document type and contract type). The Al system or hybrid Al system identifies sections of the selected documents relevant to the response (e.g., Identify sections relevant to response), composes the response (e.g., Compose response), and provides the response at block 230 (e.g., Display response).

[0225] In an embodiment, the method supports analysis of a document library comprising market data contracts, data exploitation agreements, related financial documents, or a combination thereof. The user can submit one or more queries directed to specific data rights issues, for example, whether a particular market data agreement permits redistribution of the data or the creation of derivative works, and the Al system can process the queries against the domain knowledge base to provide a plain-language response identifying the applicable provisions and their sources within the document library, with interactive references to the documents and passages relied upon.

[0226] In an embodiment, the method supports analysis of a document library comprising regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions (for example, filings directed to specific regulatory or supervisory bodies such as securities regulators, commodities regulators, or financial supervisory authorities), or a combination thereof. The user can submit one or more queries directed to specific regulatory requirements or compliance gaps applicable to the user's regulatory context, and the Al system can process the queries against the domain knowledge base to provide a plain-language response identifying the relevant obligations and the documents from which they derive, with interactive references enabling access to the supporting sources.

[0227] In an embodiment, the method can receive from the user one or more additional queries and can repeat the processing and providing steps for the one or more additional queries. This successive-query capability enables the user to submit a sequence of queries concerning different issues arising in a transaction or compliance matter and to populate cited documents, schedules, and other instruments required to complete the transaction. In an embodiment, the Al system can subsequently assist the user in amending those documents as the terms of the transaction evolve.EXAMPLESExample 1Summary

[0228] DataRightsAI (DRAI) is revolutionizing how financial firms extract critical contract data about rights and obligations, delivering instant insights for real-time, informed decision-making. As a next-generation contract analysis solution, DRAI embeds expert-curated domain knowledge, offering an unparalleled value proposition absent in today's market. This differentiation has been validated through extensive interviews with POC clients, potential customers, and market research. DRAI’s founding team includes industry-renowned technologists with proven success at scale, a visionary leader in global business expansion, and an award-winning attorney and legal innovator.The Problem

[0229] Companies must have a clear understanding of the rights and obligations embedded in their contracts. While there are many contract types and use cases that DRAI’s technology can address, DRAI is initially focused on data. Demand for data on Wall Street is insatiable. Firms increasingly seek to assess their data assets to unlock new revenue opportunities such as the creation of ETFs, indices, data streams, and beyond. The stakes of not knowing are high: lost revenue, stalled innovation, operational inefficiencies, and hidden risks. Yet, determining rights and obligations in contracts remains a major problem. Traditional contract analysis is slow, expensive, and labor-intensive, requiring deep legal expertise to parse complex language and intricate agreement structures. As a result, companies either endure costly delays or forgo lucrative opportunities with their data assets. DRAI’s technology unlocks rights and obligations in contracts in an instant.Solution and Key Functionality

[0230] Analyzes and explains in plain English a company’s rights and obligations. Identifies conflicts and risks (e.g., ownership disputes, sublicensing restrictions). Comparesamendments and historical versions to track contractual changes. Provides instant legal insights without reliance on external legal counsel. Integrates with existing contract management systems via API.

[0231] DRAI is expertly trained to analyze complex documents — contracts, policies, and regulations — with precision. Built on legal expert-driven interpretive methods, it is not a black box contextualizing, comprehending, interpreting, and delivering insights in plain English, not just legalese excerpts that have been cut and pasted.

[0232] The Process - simplified: DRAI first understands a document’s business context; DRAI then applies a structured ontology to answer questions and resolve issues accurately; DRAI complements every output with an audit trail and customizable reports tailored to multiple queries.

[0233] Quantifiable savings of Time and Money: DRAI delivers results in minutes — not billable hours over days or weeks — saving businesses time and money. Seamlessly integrating into existing workflows, it transforms productivity, scalability, and cost efficiency. One company estimated for ONE Query saving $300,000 in outside billable hours and 6 weeks of work that Datarightsai delivered the results in minutes. ( $50,000 billable hours per week x 6 weeks). Another client shared an audit on over 120 documents that cost $600,000 in outside counsel fees and three weeks for a result which DRAI was able to achieve in minutes.Value Proposition

[0234] On-demand interrogation = faster insights, iterative results.

[0235] Dynamic updates = real-time analysis accounts for contract and policy evolution.

[0236] Translatable output = seamless communication across legal, product, C-suite, board.

[0237] In-house enablement = accelerated productivity, refocused capacity to higher-value add efforts, and decreased external legal costs.

[0238] Central storage = saved historical queries and results provides one-stop-shop access to contracts, insights and an audit trail to reduce key-man risk.

[0239] Context-driven analysis & comprehension = increased accuracy, tailored and immediately actionable insights.

[0240] Primary end-users are product teams, procurement and in-house lawyers. However, C-suite and Boards should have a vested interest in DRAI adoption to accelerate monetization, reduce costs and minimize risks.Example 2Using Knowledge Curation to Improve the Accuracy of Al Agents for High-Stakes Legal Applications in Financial Services, such as Data Ownership Rights / . Introduction and Context

[0241] Recent developments in Large Language Models (LLMs) have unlocked new possibilities in automating legal analysis and decision-making. However, legal documents pose unique challenges: Complexity and Nuance, legal texts often contain intricate clauses, ambiguous language, and context-specific terminology. Many high-stakes documents, such as market data contracts, contain extensive definitions and cross-references, which require a thorough understanding of the text’s structure; Accuracy Demands, in high-stakes legal environments, even minor inaccuracies can lead to significant consequences. For instance, a misinterpretation of ownership clauses in a market data agreement could expose an organization to financial or regulatory risks; Dynamic Nature, laws and legal interpretations continuously evolve. Consequently, Al systems must adapt and update their knowledge over time to remain relevant and accurate.

[0242] Traditional Retrieval-Augmented Generation (RAG) systems, which combine document retrieval with LLM-generated responses, struggle to fully capture these nuances, often leading to errors or misinterpretations. Below is an example of such an error using the most potent reasoning model (with access to 20 PDF documents). It incorrectly infers the definition of “original works” and provides an incorrect answer with dangerously high confidence.2. Limitations of Traditional RAG Systems for Legal Documents

[0243] 1. Surface-Level Retrieval: RAG systems often rely on keyword matching and statistical relevance, which can miss deeper semantic relationships or implied conditions in legal clauses.

[0244] 2. Ontology Mismatches: Different bodies of legal documents (e.g., CME forms, NYSE contracts) use their own unique ontologies. Without a structured approach, it becomes challenging to reconcile differing terminologies and definitions across multiple sources.

[0245] 3. Loss of Structural Complexity: Legal documents typically have hierarchical clause structures, cross-references, and dedicated definition sections. Traditional RAG pipelines are not designed to preserve or interpret this level of detail, risking incomplete or incorrect analysis.

[0246] 4. Hallucination Risks: LLMs may generate plausible but factually incorrect interpretations if not guided by precise, validated context, particularly dangerous in high-stakes legal settings.3. Our Approach: Knowledge Curation

[0247] Knowledge curation is a deliberate, expert-driven process that systematically captures, organizes, and validates domain-specific information before integrating it into Al systems. This approach ensures that the Al agent has reliable, up-to-date knowledge, mitigating many of the pitfalls seen in traditional RAG methods.3.1 Overview of Knowledge Curation

[0248] 1. Identify Critical Concepts and Clauses: Experts and automated tools work together to locate key definitions, clauses, and domain-specific terminologies.

[0249] 2. Extract and Validate: Extracted information is cross-checked with legal experts to ensure accuracy. This step is particularly crucial for market data agreements, where subtle differences in wording can drastically change interpretations of ownership or usage rights.

[0250] 3. Normalize and Organize: Definitions and relationships are standardized into a consistent ontology. This step is essential for reconciling different document sources (e.g., CME, NYSE) and any overarching cross-exchange framework.

[0251] 4. Maintain a Dynamic Repository: Legal rules evolve, and so must the knowledge repository. Ongoing updates and validations keep the curated data current, reducing the risk of outdated interpretations.

[0252] Figure 2 Example of a Knowledge Graph Created by Human Domain Experts 3.2 Example Knowledge Curation Pipeline

[0253] Figure 3 provides a conceptual overview of how knowledge curation feeds into Al system to answer complex legal questions:

[0254] 1. Question Enrichment: The system refines the user’s query to align with domainspecific terminology.

[0255] 2. Document Filtering: Relevant documents are selected based on curated metadata and definitions.

[0256] 3. Chunk Extraction: Relevant sections of text (e.g., definitions, clauses) are segmented for detailed analysis.

[0257] 4. Ordering and Document Relationships: Cross-references and clause hierarchies are resolved.

[0258] 5. Synthesis: The system compiles curated information into a coherent response.

[0259] 6. Scoring: Each response is evaluated for accuracy and confidence, ensuring only reliable outputs are returned.4. Example Use Case: Market Data Ownership Contracts

[0260] Consider a scenario in which an organization must review multiple market data ownership contracts from various exchanges. Traditional RAG might: Retrieve clauses based on keyword matching (e.g., “ownership,” “IP rights”) but miss crucial context found in cross-references or definition sections; Generate incomplete summaries, failing to capture how subtle differences in defined terms could alter contractual obligations.

[0261] By contrast, a knowledge-curated system would: Identify and extract the precise definitions of “Data,” “Derived Information,” or “Proprietary Data” from each contract; Map out cross-references (e.g., references to earlier clauses, notice periods, or usage limitations); Ensure consistent interpretation across multiple sources, reconciling any differences in ontologies (e.g., CME vs. NYSE definitions).5. Conclusion and Future Directions

[0262] Knowledge curation offers a robust framework to address the inherent limitations of traditional RAG systems, especially in high-stakes legal domains. By systematically capturing and validating the nuances of legal documents, such as market data ownership contracts, this approach ensures that Al agents deliver accurate, reliable, and context-aware outputs.

[0263] Future Work: Scaling Curation Processes: Automating parts of expert validation and annotation to handle an ever-expanding corpus of legal documents; Refining Ontology Reconciliation: Further improving how differing ontologies (e.g., across multiple exchanges) are integrated into a unified knowledge base; Dynamic Updates and Real-Time Monitoring: Continuously updating curated repositories to reflect changes in legal interpretations or newly introduced regulations.Example AspectsStatement 1. An artificial intelligence (Al) system, comprising:a computing device comprising one or more processors, the computing device configured to receive multimodal queries and prompts and to provide multimodal responses in one or more human and computer languages to the queries based on reference data that can include single or multiple sources of the reference data for processing simultaneously;wherein the computing device includes memory coupled to the one or more processors, the memory configured to store reference data and curated domain knowledge including codefor execution by the one or more processors, the one or more processors further coupled to another memory configured to store the reference data and domain knowledge during a plurality of training sessions conducted by a legal professional using at least one input device coupled to the first computing device;wherein the Al system employs in one embodiment a large language model (LLM) or employs in another embodiment any other future models for interpretation of human or computer language either through a network connection to an externally hosted LLM or an internally hosted LLM or by incorporating the LLM on the computing device itself without affecting integrity and persistence of the reference data;wherein the domain knowledge is curated and applied in a highly selective manner to the reference data used by the one or more processors to respond to the queries via a user interface related to legal and regulatory documents in financial services.Statement 2. The system of Statements 1, wherein the domain knowledge comprise a knowledge graph or other representation of domain knowledge in a manner interpretable by a processor and communicated to another processor or a human.Statement s. The system of either of Statements 1 or 2, wherein the domain knowledge comprises information particular to an industry.Statement 4. The system of any of Statements 1-3, wherein the domain knowledge comprises a reference to one or more documents, key concepts relating to or included within the one or more documents, text excerpts and logical expressions interpretations from any of the one or more documents, and an indication of relationships between the agreements or the key concepts.Statement 5. The system of any of Statements 1-4, whereinthe domain knowledge comprises a plurality of hierarchical levels; andthe at least one input device is configured to receive selections of different regions or icons within the user interface to identify further information relevant to a user query.Statement 6. The system of any of Statements 1-5, wherein the domain knowledge specific to a domain is user-editable and revisable for an authorized user of the system.Statement 7. The system of any of Statements 1-6, wherein the ancillary tasks of the LLM comprise:probing a user query to further determine a nature of or intent behind the inquiry; summarizing a section of one or more documents or other text excerpts.Statement 8. The system of any of Statements 1-7, wherein a decision process comprising reasoning supporting a response to a query is under the control of the code and the reference data and the reasoning support is displayed to the user of the system to validate the system’s comprehension of the queries or prompts.Statement 9. The system of any of Statements 1-8, whereby the Al model is tuned and refined by the domain knowledge in a static or dynamic fashion.Statement 10. The system of any of Statements 1-9 wherein the responses can be formatted in a flexible reporting format that is multilingual and multimodal in response to queries or promptsStatement 11. The system of any of Statements 1-10, wherein a response to a query or a prompt can be downloaded and manipulated and transmitted to another user by a human or an automated computer interface.Statement 12. The system of any of Statements 1-11 where updates to the reference data occur and are updated in real time.Statement 13. The system of any of Statements 1-12, where the response to the query articulates a conclusion and the reasoning used by the system is displayed along with a weighting factor used by the Al system as a validation of the correctness of the response to the query or prompt.Statement 14. A method, comprising the steps performed by the Al system in any of Statements 1-13.General Definitions

[0264] Unless defined otherwise, technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0265] As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items. Expressions such as “at least one of,” when preceding a list of elements, modify the entire list of elements and do not modify the individual elements of the list.

[0266] As used herein, the singular forms “a”, “an”, and “the” include both singular and plural referents unless the context clearly dictates otherwise. As used in the specification and the appended claims, the singular forms “a,” “an” and “the” include plural referents unless the context clearly dictates otherwise. Thus, for example, reference to “a processor,” “a large language model,” or “a conduit,” includes, but is not limited to, two or more such processors, large language models, or conduits, and the like.

[0267] The term “optional” or “optionally” means that the subsequent described event, circumstance or substituent may or may not occur, and that the description includes instances where the event or circumstance occurs and instances where it does not.

[0268] The recitation of numerical ranges by endpoints includes all numbers and fractions subsumed within the respective ranges, as well as the recited endpoints.

[0269] The terms “about” or “approximately” as used herein when referring to a measurable value such as a parameter, an amount, a temporal duration, and the like, are meant to encompass variations of and from the specified value, such as variations of + / -10% or less, + / -5% or less, + / -1% or less, and + / -0.1% or less of and from the specified value, insofar such variations are appropriate to perform in the disclosed invention. It is to be understood that the value to which the modifier “about” or “approximately” refers is itself also specifically, and preferably, disclosed.

[0270] Various embodiments are described hereinafter. It should be noted that the specific embodiments are not intended as an exhaustive description or as a limitation to the broader aspects discussed herein. One aspect described in conjunction with a particular embodiment is not necessarily limited to that embodiment and can be practiced with any other embodiment(s). Reference throughout this specification to “one embodiment”, “an embodiment,” “an example embodiment,” means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrases “in one embodiment,” “in an embodiment,” or “anexample embodiment” in various places throughout this specification are not necessarily all referring to the same embodiment, but may. Furthermore, the particular features, structures or characteristics may be combined in any suitable manner, as would be apparent to a person skilled in the art from this disclosure, in one or more embodiments. Furthermore, while some embodiments described herein include some but not other features included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the invention. For example, in the appended claims, any of the claimed embodiments can be used in any combination.

[0271] All publications, published patent documents, and patent applications cited herein are hereby incorporated by reference to the same extent as though each individual publication, published patent document, or patent application was specifically and individually indicated as being incorporated by reference.

[0272] All documents mentioned herein are incorporated herein by reference to disclose and describe the methods and / or materials in connection with which the publications are cited. The publications discussed herein are provided solely for their disclosure prior to the filing date of the present application. Nothing herein is to be construed as an admission that the present invention is not entitled to antedate such publication by virtue of prior invention.

[0273] As will be apparent to those of skill in the art upon reading this disclosure, each of the individual embodiments described and illustrated herein has discrete components and features which may be readily separated from or combined with the features of any of the other several embodiments without departing from the scope or spirit of the present disclosure.

[0274] While aspects of the present disclosure can be described and claimed in a particular statutory class, such as the system statutory class, this is for convenience only and one of skill in the art will understand that each aspect of the present disclosure can be described and claimed in any statutory class.

[0275] It is also to be understood that the terminology used herein is for the purpose of describing particular aspects only and is not intended to be limiting. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the disclosed compositions and methods belong. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the specification and relevant art and should not be interpreted in an idealized or overly formal sense unless expressly defined.

[0276] Prior to describing the various aspects of the present disclosure, the following definitions are provided and should be used unless otherwise indicated. Additional terms may be defined elsewhere in the present disclosure.

[0277] Further embodiments are illustrated in the following Examples which are given for illustrative purposes only and are not intended to limit the scope of the invention.***

[0278] Various modifications and variations of the described methods, pharmaceutical compositions, and kits of the invention will be apparent to those skilled in the art without departing from the scope and spirit of the invention. Although the invention has been described in connection with specific embodiments, it will be understood that it is capable of further modifications and that the invention as claimed should not be unduly limited to such specific embodiments. Indeed, various modifications of the described modes for carrying out the invention that are obvious to those skilled in the art are intended to be within the scope of the invention. This application is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure come within known customary practice within the art to which the invention pertains and may be applied to the essential features herein before set forth.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for analyzing documents, comprising:a. receiving, from a user, one or more queries associated with a document library comprising a plurality of structured domain documents;b. processing, by an Al system or hybrid Al system, the one or more queries using a domain knowledge base curated by one or more domain experts; andc. providing, to the user, a response to the one or more queries, wherein the response is in plain language and comprises a reference identifying one or more of the documents relied upon to generate the response.

2. The method of claim 1, wherein the Al system or hybrid Al system comprises one or more machine learning models comprising one or more of a large language model (LLM), a small language model (SLM), a topic model, a Latent Dirichlet Allocation (LDA) model, a neural network, or a combination thereof.

3. The method of any of the preceding claims, wherein the Al system or hybrid Al system comprises a small language model (SLM).

4. The method of any of the preceding claims, wherein the reference comprises an interactive reference to one or more of the documents, one or more passages within the one or more documents relied upon to generate the response, or a combination thereof.

5. The method of any of the preceding claims, wherein processing the one or more queries comprises transforming the one or more queries into one or more enriched queries by aligning terminology of the one or more queries with domain-specific language of the domain knowledge base.

6. The method of any of the preceding claims, wherein the one or more queries comprise free-form queries, queries selected from a domain-specific query library associated with the domain knowledge base, or a combination thereof.

7. The method of any of the preceding claims, wherein the domain knowledge base comprises a knowledge graph or other symbolic representation of knowledge comprising a plurality of nodes and a plurality of edges.

8. The method of any of the preceding claims, wherein the knowledge graph or other symbolic representation of knowledge is a network graph.

9. The method of any of the preceding claims, wherein the network graph comprises a hierarchical graph, a directed acyclic graph, a property graph, an ontology graph, a semantic network graph, a multi-relational graph, a hybrid thereof, or a combination thereof.

10. The method of any of the preceding claims, wherein the plurality of nodes represent one or more document types, agreement provisions, statutes, statutory provisions, regulatory frameworks, compliance requirements, risk factors, jurisdictions, entities, concepts, or a combination thereof.

11. The method of any of the preceding claims, wherein the plurality of edges represent relationships between the nodes.

12. The method of any one of claims 7-11, wherein the plurality of edges independently comprise one or more of a supersession relationship, a section relationship, a read-for relationship, a related-to relationship, a jurisdictional relationship, a compliance relationship, a risk relationship, a cross-reference relationship, a level relationship, or a combination thereof.

13. The method of any one of claims 7-12, wherein the processing comprises traversing the knowledge graph or other symbolic representation of knowledge to retrieve one or more portions of the domain knowledge base relevant to the one or more queries.

14. The method of claim 13, wherein traversing the knowledge graph or other symbolic representation of knowledge comprises traversing from an organizational level to a section level to a clause level to a context level.

15. The method of any of the preceding claims, further comprising receiving, from the user, a modification to one or more nodes or edges of the knowledge graph or other symbolic representation of knowledge, and updating the knowledge graph or other symbolic representation of knowledge to reflect the modification.

16. The method of any of the preceding claims, wherein the processing is deterministic.

17. The method of claim 16, wherein the method further comprises generating a reasoning chain tracing each step taken to generate the response.

18. The method of claim 16, wherein the method further comprises providing, to the user, an indication of one or more processing steps being performed during the processing.

19. The method of any of the preceding claims, wherein processing the one or more queries comprises processing the one or more queries simultaneously against a plurality of documents in the document library.

20. The method of claim 19, wherein the plurality of documents comprises at least 10 documents, at least 100 documents, at least 1,000 documents, at least 10,000 documents, or at least 100,000 documents.

21. The method of claim 19, wherein the response comprises a matrix, a summary of the response, a visualization of the response, or a combination thereof.

22. The method of claim 21, wherein the matrix comprises a plurality of cells, each cell corresponding to one or more documents of the plurality of documents.

23. The method of claim 22, wherein each cell of the matrix comprises a plain-language response for the one or more documents corresponding to that cell.

24. The method of claim 22, wherein each cell of the matrix comprises an identification of one or more passages within the one or more documents corresponding to that cell relied upon to generate the response for that cell.

25. The method of any one of claims 22-24, wherein rows and columns of the matrix are transposable.

26. The method of any one of claims 22-24, wherein one or more rows of the matrix each correspond to one or more of the one or more queries, and one or more columns of the matrix each correspond to an organizational level, a section level, a clause level, a context level, or a combination thereof.

27. The method of any one of claims 22-26, wherein each cell of the matrix is individually inspectable to display one or more source documents and a reasoning chain underlying the response for that cell.

28. The method of any of the preceding claims, wherein the response further comprises score types associated with each of the one or more passages relied upon to generate the response.

29. The method of any one of claims 21-28, wherein the visualization of the response comprises a domain-specific visualization corresponding to a type of domain associated with the one or more queries.

30. The method of any one of claims 21-29, wherein the visualization of the response comprises one or more nodes and edges of the knowledge graph or other symbolic representation of knowledge traversed during the processing.

31. The method of any of the preceding claims, further comprising, when the domain knowledge base does not contain information responsive to the one or more queries, returning an unsure response rather than generating an answer.

32. The method of any of the preceding claims, further comprising displaying a document viewer.

33. The method of claim 32, wherein the document viewer displays one or more documents, one or more passages within the one or more documents, or a combination thereof relied upon to generate the response.

34. The method of claim 33, wherein the one or more passages are highlighted within the document viewer.

35. The method of any of the preceding claims, wherein the domain knowledge comprises domain-specific terminology for one or more document types, cross-reference mappings between related agreement sections, document priority hierarchies, or a combination thereof.

36. The method of any of the preceding claims, wherein the document library is stored in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users.

37. The method of any of the preceding claims, wherein the processing comprises:i. aligning query terminology with domain-specific language using an ontology and definitions module;ii. refining the one or more queries against the domain knowledge base using a question enrichment module;iii. selecting relevant documents from the document library using a document filtering module;iv. traversing the domain knowledge base from an organizational level to a section level to a clause level to a context level to identify relevant portions of the domain knowledge base using a traversal module;v. segmenting the relevant portions into text segments using a chunk extraction module;vi. resolving clause hierarchies and cross-references using an ordering and document relationships module;vii. compiling and composing a coherent response using a synthesis module; viii. validating logical soundness of the compiled response against the domain knowledge base using a reasoning module;ix. evaluating relevance of domain knowledge relied upon to generate the response using a scoring module;x. or a combination thereof.

38. The method of claim 37, wherein one or more of the ontology and definitions module, the question enrichment module, the document filtering module, the traversal module, the chunk extraction module, the ordering and document relationships module, the synthesis module, the reasoning module, or the scoring module operate according to a document-type-specific configuration template that governs the order and manner in which the one or more modules are invoked for a given document type.

39. The method of any of the preceding claims, further comprising receiving, from the user, one or more additional queries and repeating the processing and providing steps for the one or more additional queries.

40. The method of any of the preceding claims, wherein the structured domain documents comprise contracts, legal documents, statutes, regulatory documents, financial documents, compliance documents, or a combination thereof.

41. The method of claim 40, wherein the contracts, legal documents, the financial documents, or a combination thereof comprise commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, government contracts, securities filings, loan agreements, or a combination thereof.

42. The method of claim 40, wherein the regulatory documents, the compliance documents, or a combination thereof comprise regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions, or a combination thereof.

43. The method of any of the preceding claims, further comprising storing the response on a local storage device, a cloud-based storage service, or a combination thereof.

44. The method of any of the preceding claims, further comprising transmitting the response to another user, a computing system, or a combination thereof.

45. The method of any of the preceding claims, wherein the domain knowledge base is generated during a plurality of training sessions conducted by one or more domain experts using data associated with the user, domain-specific curated data, or a combination thereof.

46. A method of generating a domain knowledge base for use in analyzing queries associated with documents, comprising:a. assembling reference data comprising one or more documents and document-type-specific information associated with the one or more documents; andb. curating domain knowledge from the reference data during one or more training sessions conducted by one or more domain experts to generate a domain knowledge base.

47. The method of claim 46, wherein curating the domain knowledge comprises generating a knowledge graph or other symbolic representation of knowledge comprising a plurality of nodes representing one or more document types, agreement provisions, statutory provisions, regulatory frameworks, compliance requirements, risk factors, jurisdictions, entities, concepts, or a combination thereof, and a plurality of edges representing relationships between the nodes.

48. The method of claim 47, wherein the plurality of edges independently comprise one or more of a supersession relationship, a section relationship, a read-for relationship, a related-to relationship, a jurisdictional relationship, a compliance relationship, a risk relationship, a cross-reference relationship, a level relationship, or a combination thereof.

49. The method of any one of claims 46-48, wherein curating the domain knowledge comprises encoding document priority hierarchies that define an order of precedence among two or more of the one or more documents.

50. The method of any one of claims 46-49, wherein curating the domain knowledge comprises disambiguating domain-specific terminology by associating one or more terms with context-specific meanings based on the document type in which each term appears.

51. The method of any one of claims 46-50, wherein curating the domain knowledge comprises mapping cross-references between related agreement sections across two or more of the one or more documents.

52. The method of any one of claims 46-51, further comprising conducting one or more additional training sessions to extend or update the domain knowledge base following an initial set of training sessions.

53. The method of any one of claims 46-52, further comprising generating a document-type-specific configuration template for one or more document types, the document-type-specific configuration template governing the order and manner in which one or more processing modules are invoked when analyzing a document of that document type.

54. The method of any one of claims 46-53, wherein the one or more domain experts have expertise in commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, statutes, regulatory filings, compliance policies, merger and acquisition due diligence, taxation, government contracts, securities filings, loan agreements, statutory provisions, internal controls, government regulatory submissions, or a combination thereof.

55. The method of any one of claims 46-54, wherein the reference data comprises contracts, legal documents, regulatory documents, financial documents, compliance documents, or a combination thereof, comprising one or more commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, government contracts, securities filings, loan agreements, regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions, or a combination thereof.

56. The method of any one of claims 46-55, wherein the domain knowledge base is structured to support deterministic, rule-based processing of queries.

57. The method of any one of claims 46-56, further comprising storing the domain knowledge base in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users.

58. The method of any one of claims 46-57, further comprising establishing one or more scoring parameters or relevance thresholds for evaluating domain knowledge relied upon during processing of queries against the domain knowledge base.

59. The method of any one of claims 46-58, wherein the reference data comprises data associated with the user, domain-specific curated data, or a combination thereof.

60. A system to analyze documents, comprising:a storage device; anda processor communicatively coupled to the storage device, wherein the processor executes application code instructions that are stored in the storage device to cause the system to:a. receive, by one or more computing devices, one or more queries from a user, the one or more queries associated with a document library comprising a plurality of structured domain documents;b. process, by an Al system or hybrid Al system, the one or more queries using a domain knowledge base curated by one or more domain experts; andc. transmit, by the one or more computing devices, a response to the one or more queries to a device associated with the user, wherein the response is in plain language and comprises a reference identifying one or more of the documents relied upon to generate the response.

61. The system of claim 60, wherein the Al system or hybrid Al system comprises one or more machine learning models comprising one or more of a large language model (LLM), a small language model (SLM), a topic model, a Latent Dirichlet Allocation (LDA) model, a neural network, or a combination thereof.

62. The system of any one of claims 60-61, wherein the Al system or hybrid Al system comprises a small language model (SLM).

63. The system of any one of claims 60-62, wherein the reference comprises an interactive reference to one or more of the documents, one or more passages within the one or more documents relied upon to generate the response, or a combination thereof.

64. The system of any one of claims 60-63, wherein processing the one or more queries comprises transforming the one or more queries into one or more enriched queries by aligning terminology of the one or more queries with domain-specific language of the domain knowledge base.

65. The system of any one of claims 60-64, wherein the one or more queries comprise free-form queries, queries selected from a domain-specific query library associated with the domain knowledge base, or a combination thereof.

66. The system of any one of claims 60-65, wherein the domain knowledge base comprises a knowledge graph or other symbolic representation of knowledge comprising a plurality of nodes and a plurality of edges.

67. The system of any one of claims 60-66, wherein the knowledge graph or other symbolic representation of knowledge is a network graph.

68. The system of any one of claims 60-67, wherein the network graph comprises a hierarchical graph, a directed acyclic graph, a property graph, an ontology graph, a semantic network graph, a multi-relational graph, a hybrid thereof, or a combination thereof.

69. The system of any one of claims 60-68, wherein the plurality of nodes represent one or more document types, agreement provisions, statutes, statutory provisions, regulatory frameworks, compliance requirements, risk factors, jurisdictions, entities, concepts, or a combination thereof.

70. The system of any one of claims 60-69, wherein the plurality of edges represent relationships between the nodes.

71. The system of any one of claims 66-70, wherein the plurality of edges independently comprise one or more of a supersession relationship, a section relationship, aread-for relationship, a related-to relationship, a jurisdictional relationship, a compliance relationship, a risk relationship, a cross-reference relationship, a level relationship, or a combination thereof.

72. The system of any one of claims 66-71, wherein the processing comprises traversing the knowledge graph or other symbolic representation of knowledge to retrieve one or more portions of the domain knowledge base relevant to the one or more queries.

73. The system of claim 72, wherein traversing the knowledge graph or other symbolic representation of knowledge comprises traversing from an organizational level to a section level to a clause level to a context level.

74. The system of any one of claims 60-73, wherein the application code instructions further cause the system to receive, from the user, a modification to one or more nodes or edges of the knowledge graph or other symbolic representation of knowledge, and update the knowledge graph or other symbolic representation of knowledge to reflect the modification.

75. The system of any one of claims 60-74, wherein the processing is deterministic.

76. The system of claim 75, wherein the application code instructions further cause the system to generate a reasoning chain tracing each step taken to generate the response.

77. The system of claim 75, wherein the application code instructions further cause the system to transmit, to a device associated with the user, an indication of one or more processing steps being performed during the processing.

78. The system of any one of claims 60-77, wherein processing the one or more queries comprises processing the one or more queries simultaneously against a plurality of documents in the document library.

79. The system of claim 78, wherein the plurality of documents comprises at least 10 documents, at least 100 documents, at least 1,000 documents, at least 10,000 documents, or at least 100,000 documents80. The system of claim 78, wherein the response comprises a matrix, a summary of the response, a visualization of the response, or a combination thereof.

81. The system of claim 80, wherein the matrix comprises a plurality of cells, each cell corresponding to one or more documents of the plurality of documents.

82. The system of claim 81, wherein each cell of the matrix comprises a plain-language response for the one or more documents corresponding to that cell.

83. The system of claim 81, wherein each cell of the matrix comprises an identification of one or more passages within the one or more documents corresponding to that cell relied upon to generate the response for that cell.

84. The system of any one of claims 81-83, wherein rows and columns of the matrix are transposable.

85. The system of any one of claims 81-83, wherein one or more rows of the matrix each correspond to one or more of the one or more queries, and one or more columns of the matrix each correspond to an organizational level, a section level, a clause level, a context level, or a combination thereof.

86. The system of any one of claims 81-85, wherein each cell of the matrix is individually inspectable to display one or more source documents and a reasoning chain underlying the response for that cell.

87. The system of any one of claims 60-86, wherein the response further comprises score types associated with each of the one or more passages relied upon to generate the response.

88. The system of any one of claims 80-87, wherein the visualization of the response comprises a domain-specific visualization corresponding to a type of domain associated with the one or more queries.

89. The system of any one of claims 80-88, wherein the visualization of the response comprises one or more nodes and edges of the knowledge graph or other symbolic representation of knowledge traversed during the processing.

90. The system of any one of claims 60-89, wherein the application code instructions further cause the system to, when the domain knowledge base does not contain information responsive to the one or more queries, return an unsure response rather than generating an answer.

91. The system of any one of claims 60-90, wherein the application code instructions further cause the system to display a document viewer.

92. The system of claim 91, wherein the document viewer displays one or more documents, one or more passages within the one or more documents, or a combination thereof relied upon to generate the response.

93. The system of claim 92, wherein the one or more passages are highlighted within the document viewer.

94. The system of any one of claims 60-93, wherein the domain knowledge comprises domain-specific terminology for one or more document types, cross-reference mappings between related agreement sections, document priority hierarchies, or a combination thereof.

95. The system of any one of claims 60-94, wherein the document library is stored in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users.

96. The system of any one of claims 60-95, wherein the processing comprises:i. aligning query terminology with domain-specific language using an ontology and definitions module;ii. refining the one or more queries against the domain knowledge base using a question enrichment module;iii. selecting relevant documents from the document library using a document filtering module;iv. traversing the domain knowledge base from an organizational level to a section level to a clause level to a context level to identify relevant portions of the domain knowledge base using a traversal module;v. segmenting the relevant portions into text segments using a chunk extraction module;vi. resolving clause hierarchies and cross-references using an ordering and document relationships module;vii. compiling and composing a coherent response using a synthesis module; viii. validating logical soundness of the compiled response against the domain knowledge base using a reasoning module;ix. evaluating relevance of domain knowledge relied upon to generate the response using a scoring module;x. or a combination thereof.

97. The system of claim 96, wherein one or more of the ontology and definitions module, the question enrichment module, the document filtering module, the traversal module, the chunk extraction module, the ordering and document relationships module, the synthesis module, the reasoning module, or the scoring module operate according to a document-type-specific configuration template that governs the order and manner in which the one or more modules are invoked for a given document type.

98. The system of any one of claims 60-97, wherein the application code instructions further cause the system to receive, from the user, one or more additional queries and repeat the processing and transmission steps for the one or more additional queries.

99. The system of any one of claims 60-98, wherein the structured domain documents comprise contracts, legal documents, statutes, regulatory documents, financial documents, compliance documents, or a combination thereof.

100. The system of claim 99, wherein the contracts, the legal documents, the financial documents, or a combination thereof comprise commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, government contracts, securities filings, loan agreements, or a combination thereof.

101. The system of claim 100, wherein the regulatory documents, the compliance documents, or a combination thereof comprise regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions, or a combination thereof.

102. The system of any one of claims 60-101, wherein the application code instructions further cause the system to store the response on a local storage device, a cloud-based storage service, or a combination thereof.

103. The system of any one of claims 60-102, wherein the application code instructions further cause the system to transmit the response to another user, a computing system, or a combination thereof.

104. A computer program product, compri sing :a non-transitory computer-readable storage device having computer-executable program instructions embodied thereon that when executed by a computer cause the computer to analyze documents, the computer-executable program instructions comprising:a. computer-executable program instructions to receive, by one or more computing devices, one or more queries from a user, the one or more queries associated with a document library comprising a plurality of structured domain documents;b. computer-executable program instructions to process, by an Al system or hybrid Al system, the one or more queries using a domain knowledge base curated by one or more domain experts; andc. computer-executable program instructions to transmit, by the one or more computing devices, a response to the one or more queries to a device associated with the user, wherein the response is in plain language and comprises a reference identifying one or more of the documents relied upon to generate the response.

105. The computer program product of claim 104, wherein the Al system or hybrid Al system comprises one or more machine learning models comprising one or more of a large language model (LLM), a small language model (SLM), a topic model, a Latent Dirichlet Allocation (LDA) model, a neural network, or a combination thereof.

106. The computer program product of any one of claims 104-105, wherein the Al system or hybrid Al system comprises a small language model (SLM).

107. The computer program product of any one of claims 104-106, wherein the reference comprises an interactive reference to one or more of the documents, one or more passages within the one or more documents relied upon to generate the response, or a combination thereof.

108. The computer program product of any one of claims 104-107, wherein processing the one or more queries comprises transforming the one or more queries into one or more enriched queries by aligning terminology of the one or more queries with domain-specific language of the domain knowledge base.

109. The computer program product of any one of claims 104-108, wherein the one or more queries comprise free-form queries, queries selected from a domain-specific query library associated with the domain knowledge base, or a combination thereof.

110. The computer program product of any one of claims 104-109, wherein the domain knowledge base comprises a knowledge graph or other symbolic representation of knowledge comprising a plurality of nodes and a plurality of edges.

111. The computer program product of any one of claims 104-110, wherein the knowledge graph or other symbolic representation of knowledge is a network graph.

112. The computer program product of any one of claims 104-111, wherein the network graph comprises a hierarchical graph, a directed acyclic graph, a property graph, an ontology graph, a semantic network graph, a multi-relational graph, a hybrid thereof, or a combination thereof.

113. The computer program product of any one of claims 104-112, wherein the plurality of nodes represent one or more document types, agreement provisions, statutes, statutory provisions, regulatory frameworks, compliance requirements, risk factors, jurisdictions, entities, concepts, or a combination thereof.

114. The computer program product of any one of claims 104-113, wherein the plurality of edges represent relationships between the nodes.

115. The computer program product of any one of claims 110-114, wherein the plurality of edges independently comprise one or more of a supersession relationship, a section relationship, a read-for relationship, a related-to relationship, a jurisdictional relationship, a compliance relationship, a risk relationship, a cross-reference relationship, a level relationship, or a combination thereof.

116. The computer program product of any one of claims 110-115, wherein the processing comprises traversing the knowledge graph or other symbolic representation of knowledge to retrieve one or more portions of the domain knowledge base relevant to the one or more queries.

117. The computer program product of claim 116, wherein traversing the knowledge graph or other symbolic representation of knowledge comprises traversing from an organizational level to a section level to a clause level to a context level.

118. The computer program product of any one of claims 104-117, further comprising computer-executable program instructions to receive, by the one or more computing devices, a modification from the user to one or more nodes or edges of the knowledge graph or other symbolic representation of knowledge, and update the knowledge graph or other symbolic representation of knowledge to reflect the modification.

119. The computer program product of any one of claims 104-118, wherein the processing is deterministic.

120. The computer program product of claim 119, further comprising computerexecutable program instructions to generate a reasoning chain tracing each step taken to generate the response.

121. The computer program product of claim 119, further comprising computerexecutable program instructions to transmit, by the one or more computing devices, an indication of one or more processing steps being performed during the processing to a device associated with the user.

122. The computer program product of any one of claims 104-121, wherein the computer-executable program instructions to process comprise computer-executable program instructions to process the one or more queries simultaneously against a plurality of documents in the document library.

123. The computer program product of claim 122, wherein the plurality of documents comprises at least 10 documents, at least 100 documents, at least 1,000 documents, at least 10,000 documents, or at least 100,000 documents.

124. The computer program product of claim 122, wherein the response comprises a matrix, a summary of the response, a visualization of the response, or a combination thereof.

125. The computer program product of claim 124, wherein the matrix comprises a plurality of cells, each cell corresponding to one or more documents of the plurality of documents.

126. The computer program product of claim 125, wherein each cell of the matrix comprises a plain-language response for the one or more documents corresponding to that cell.

127. The computer program product of claim 125, wherein each cell of the matrix comprises an identification of one or more passages within the one or more documents corresponding to that cell relied upon to generate the response for that cell.

128. The computer program product of any one of claims 125-127, wherein rows and columns of the matrix are transposable.

129. The computer program product of any one of claims 125-127, wherein one or more rows of the matrix each correspond to one or more of the one or more queries, and one or more columns of the matrix each correspond to an organizational level, a section level, a clause level, a context level, or a combination thereof.

130. The computer program product of any one of claims 125-129, wherein each cell of the matrix is individually inspectable to display one or more source documents and a reasoning chain underlying the response for that cell.

131. The computer program product of any one of claims 104-130, wherein the response further comprises score types associated with each of the one or more passages relied upon to generate the response.

132. The computer program product of any one of claims 124-131, wherein the visualization of the response comprises a domain-specific visualization corresponding to a type of domain associated with the one or more queries.

133. The computer program product of any one of claims 124-132, wherein the visualization of the response comprises one or more nodes and edges of the knowledge graph or other symbolic representation of knowledge traversed during the processing.

134. The computer program product of any one of claims 104-133, further comprising computer-executable program instructions to, when the domain knowledge base does not contain information responsive to the one or more queries, return an unsure response rather than generating an answer.

135. The computer program product of any one of claims 104-134, further comprising computer-executable program instructions to display a document viewer.

136. The computer program product of claim 135, wherein the document viewer displays one or more documents, one or more passages within the one or more documents, or a combination thereof relied upon to generate the response.

137. The computer program product of claim 136, wherein the one or more passages are highlighted within the document viewer.

138. The computer program product of any one of claims 104-137, wherein the domain knowledge comprises domain-specific terminology for one or more document types, crossreference mappings between related agreement sections, document priority hierarchies, or a combination thereof.

139. The computer program product of any one of claims 104-138, wherein the document library is stored in a dedicated partition associated with the user, the dedicated partition being inaccessible to other users or Al systems associated with other users.

140. The computer program product of any one of claims 104-139, wherein the processing comprises:i. aligning query terminology with domain-specific language using an ontology and definitions module;ii. refining the one or more queries against the domain knowledge base using a question enrichment module;iii. selecting relevant documents from the document library using a document filtering module;iv. traversing the domain knowledge base from an organizational level to a section level to a clause level to a context level to identify relevant portions of the domain knowledge base using a traversal module;v. segmenting the relevant portions into text segments using a chunk extraction module;vi. resolving clause hierarchies and cross-references using an ordering and document relationships module;vii. compiling and composing a coherent response using a synthesis module; viii. validating logical soundness of the compiled response against the domain knowledge base using a reasoning module;ix. evaluating relevance of domain knowledge relied upon to generate the response using a scoring module;x. or a combination thereof.

141. The computer program product of claim 140, wherein one or more of the ontology and definitions module, the question enrichment module, the document filtering module, the traversal module, the chunk extraction module, the ordering and document relationships module, the synthesis module, the reasoning module, or the scoring module operate according to a document-type-specific configuration template that governs the order and manner in which the one or more modules are invoked for a given document type.

142. The computer program product of any one of claims 104-141, further comprising computer-executable program instructions to receive, by the one or more computing devices, one or more additional queries from the user and repeat the processing and transmission steps for the one or more additional queries.

143. The computer program product of any one of claims 104-142, wherein the structured domain documents comprise contracts, legal documents, statutes, regulatory documents, financial documents, compliance documents, or a combination thereof.

144. The computer program product of claim 143, wherein the contracts, the legal documents, the financial documents, or a combination thereof comprise commercial vendor contracts, customer subscriber agreements, service agreements, commercial data exploitation agreements, merger and acquisition due diligence documents, taxation documents, government contracts, securities filings, loan agreements, or a combination thereof.

145. The computer program product of claim 144, wherein the regulatory documents, the compliance documents, or a combination thereof comprise regulatory filings, compliance policies, statutory provisions, internal controls documentation, government regulatory submissions, or a combination thereof.

146. The computer program product of any one of claims 104-145, further comprising computer-executable program instructions to store the response on a local storage device, a cloud-based storage service, or a combination thereof.

147. The computer program product of any one of claims 104-146, further comprising computer-executable program instructions to transmit, by the one or more computing devices, the response to another user, a computing system, or a combination thereof.