Modularized knowledge graph and retrieval enhanced large model fusion interaction method and system oriented to financial branch mechanism

Abstract

The invention discloses a financial branch mechanism-oriented modular knowledge graph and retrieval enhanced large model fusion interaction method and system. The method comprises the steps that a query vector of a target object is obtained, and the query vector comprises a problem vector used for expressing the query requirement of the target object and a post vector used for reflecting post information of the target object in a financial branch; determining a system module which can be accessed by the target object according to the post vector; determining a sub-graph region corresponding to the problem vector in the knowledge graph; and retrieving a document paragraph corresponding to the sub-graph region in a sub-index database corresponding to the system module, and inputting the document paragraph into the large model to generate a target answer corresponding to the query vector. According to the method and the device, the technical problem that a question-answering system adopted in the related technology is lack of identification of user post identities and cannot provide matched answers according to different post responsibilities and permissions of financial branches is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a method and system for integrating and interacting with a modular knowledge graph and a retrieval-enhanced large model for financial branches. Background Art

[0002] Currently, the application of large pre-trained language models (LLMs) in financial institutions faces a series of challenges, particularly when building intelligent question-answering, knowledge search, and semantic retrieval systems. Their performance and functionality are limited in various ways. For example, the question-answering systems used in these technologies lack recognition of user roles, making it impossible to provide answers tailored to the responsibilities and permissions of different positions within financial institutions. Furthermore, they struggle to ensure compliance with regulatory regulations, leading to the risk of false positives in answers generated within complex institutional systems.

[0003] To address the above-mentioned problems, no effective solutions have been proposed so far. Summary of the Invention

[0004] The embodiments of the present application provide a modular knowledge graph and retrieval-enhanced large model fusion interaction method and system for financial branches, so as to at least solve the technical problem that the question-answering system adopted by the relevant technology lacks recognition of the user's job identity and cannot provide matching answers according to the different job responsibilities and authorities of the financial branches.

[0005] According to one aspect of an embodiment of the present application, a method for integrating and interacting a modular knowledge graph and a retrieval-enhanced big model for financial branches is provided, comprising: obtaining a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirements of the target object and a position vector for reflecting the position information of the target object in the financial branch; determining a system module accessible to the target object based on the position vector, wherein the system module is a unit for logically dividing system documents according to the business areas of the financial branch; determining a subgraph area corresponding to the question vector in the knowledge graph, wherein the knowledge graph is used to represent the logical structure between system documents; retrieving a document paragraph corresponding to the subgraph area in a sub-index library corresponding to the system module, and inputting the document paragraph into the big model to generate a target answer corresponding to the query vector.

[0006] In some embodiments of the present application, determining the system modules accessible to the target object based on the position vector includes: obtaining multidimensional labels corresponding to all system modules, wherein the multidimensional labels include business domain labels for limiting the scope of application of the system document, line attribute labels for refining the business lines under the business domain, and position applicability labels for restricting access rights; determining the similarity between the position vector and the multidimensional label, and comparing the position authority information in the position vector with the position applicability label to obtain a comparison result; determining the system modules accessible to the target object based on the similarity and the comparison result, wherein each system module corresponds to a sub-index library, and the sub-index library is used to store the system documents corresponding to the system module.

[0007] In some embodiments of the present application, a knowledge graph is constructed in the following manner: extracting entities from a collection of institutional documents, wherein the entities include institutional clauses; determining the relationships between the entities, and determining the edges of the knowledge graph based on the relationships, wherein the relationships include one of the following: reference relationships, applicability relationships, inheritance relationships, and substitution relationships between institutional clauses; and constructing a knowledge graph based on the entities and edges.

[0008] In some embodiments of the present application, determining a subgraph area corresponding to a question vector in a knowledge graph includes: matching the question vector with the knowledge graph to obtain a first graph node, wherein the first graph node includes a node whose similarity with the semantic features of the question vector is greater than or equal to a first preset threshold; searching for a graph path starting from the first graph node based on the edges of the knowledge graph, and stopping the search when a stopping condition is met to obtain a second graph node set, wherein the stopping condition includes that the similarity between the semantic features of the second graph node and the question vector is less than a second preset threshold, and the second preset threshold is less than the first preset threshold; determining the subgraph area based on the first graph node and the second graph set.

[0009] In some embodiments of the present application, a document paragraph is input into a large model to generate a target answer corresponding to a query vector, including: obtaining guidance information of the large model, wherein the guidance information includes path information of the document paragraph in the knowledge graph and position information of the target object, and the path information is used to indicate the institutional logic and structural order followed by the large model when generating the answer; constructing a target template based on the guidance information, wherein the target template includes control instructions corresponding to the path information and restriction conditions corresponding to the position information; and using the large model to process the control instructions and restriction conditions to obtain the target answer.

[0010] In some embodiments of the present application, after retrieving the document paragraph corresponding to the sub-graph area in the sub-index library corresponding to the system module, the method also includes: obtaining the metadata tag corresponding to the sub-node of the sub-graph area, wherein the metadata tag includes the effective time, abolition time and applicable region of the system document; in the case where there are multiple versions of the system document corresponding to the sub-node, determining the matching score between the system document corresponding to each version and the query vector based on the metadata tag, wherein the matching score is used to quantify the applicability of the document paragraph corresponding to each version; and determining the target document paragraph for generating the target answer based on the matching score.

[0011] In some embodiments of the present application, the matching score between the institutional document corresponding to each version and the query vector is determined based on the metadata tag, including: obtaining the query time corresponding to the query vector, and determining a first indicator based on the query time, the effective time and the abolition time corresponding to the version, wherein the first indicator is used to quantify the timeliness of the version; obtaining the query area corresponding to the query vector, and determining a second indicator based on the query area and the applicable area of the version, wherein the second indicator is used to quantify the regional compliance of the version; determining a third indicator between the question vector in the query vector and the institutional document of the version, wherein the third indicator includes semantic similarity; determining the matching score based on the first indicator, the second indicator and the third indicator.

[0012] According to another aspect of an embodiment of the present application, a modular knowledge graph and retrieval-enhanced big model fusion interaction system for financial branches is also provided, including an interactive device and a server, wherein the interactive device is connected to the server and is used to obtain a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirements of the target object and a position vector for reflecting the position information of the target object in the financial branch; the server is connected to the interactive device and is used to determine the institutional module accessible to the target object based on the position vector, wherein the institutional module is a unit for logically dividing institutional documents according to the business areas of the financial branch; a subgraph area corresponding to the question vector is determined in the knowledge graph, wherein the knowledge graph is used to represent the logical structure between institutional documents; a document paragraph corresponding to the subgraph area is retrieved in the sub-index library corresponding to the institutional module, and the document paragraph is input into the big model to generate a target answer corresponding to the query vector.

[0013] According to another aspect of the embodiment of the present application, a modular knowledge graph and retrieval-enhanced big model fusion interaction device for financial branches is also provided, including: an acquisition module for acquiring a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirements of the target object and a position vector for reflecting the position information of the target object in the financial branch; a determination module for determining the institutional module accessible to the target object based on the position vector, wherein the institutional module is a unit for logically dividing institutional documents according to the business areas of the financial branch; a matching module for determining a subgraph area corresponding to the question vector in the knowledge graph, wherein the knowledge graph is used to represent the logical structure between institutional documents; an interaction module for retrieving a document paragraph corresponding to the subgraph area in the sub-index library corresponding to the institutional module, and inputting the document paragraph into the big model to generate a target answer corresponding to the query vector.

[0014] According to another aspect of the embodiment of the present application, an electronic device is also provided, including: a memory and a processor, the memory being used to store program instructions; the processor being connected to the memory and being used to execute the above-mentioned modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches.

[0015] According to another aspect of an embodiment of the present application, a non-volatile storage medium is also provided, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-mentioned modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches by running the computer program.

[0016] According to another aspect of the embodiments of the present application, a computer program product is also provided, including computer instructions, which, when executed by a processor, implement the above-mentioned modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches.

[0017] In an embodiment of the present application, a method of integrating job information and query requirements is adopted. By generating a query vector containing a question vector and a job vector, the system module that matches the target object position is screened out, and the subgraph area corresponding to the question vector is located in the knowledge graph. Then, in the sub-index library corresponding to the selected system module, a refined document paragraph retrieval is performed, and these paragraphs are input into the large model as context to generate answers, thereby achieving the purpose of accurately matching user queries with accessible system modules and accurately defining the logical boundaries of semantic retrieval, thereby achieving the technical effect of generating enhanced job adaptability and system compliance of question and answer content, improving retrieval efficiency and answer accuracy, and thus solving the technical problem that the question and answer system adopted by the related technology lacks recognition of user job identity and cannot provide matching answers according to the different job responsibilities and authorities of financial branches. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 This is a hardware structure block diagram of a computer terminal for a modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches according to an embodiment of the present application;

[0020] Figure 2 This is a flow chart of a method for integrating and interacting a modular knowledge graph with a retrieval-enhanced large model for financial branches according to an embodiment of the present application;

[0021] Figure 3 This is a structural diagram of a modular knowledge graph and retrieval-enhanced large model fusion interaction device for financial branches according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0023] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] In order to better understand the embodiments of the present application, the technical terms involved in the embodiments of the present application are explained as follows:

[0025] Large Language Model (LLM): Also known as a large model, this is a type of ultra-large-scale pre-trained language model built on deep learning technology, particularly the Transformer architecture. It aims to capture language patterns from massive amounts of text data through unsupervised learning, thereby enabling advanced natural language processing capabilities such as semantic understanding, text generation, and conversational interaction. In this embodiment of the present application, the LLM serves as the core generation engine, responsible for generating high-quality, personalized responses that meet user needs and identity background after receiving specific question vectors and job vectors.

[0026] Retrieval-Augmented Generation (RAG): A question-answering system architecture that combines information retrieval techniques with pre-trained generative models. RAG first uses semantic similarity to retrieve document fragments related to the question from a knowledge base. These fragments are then fed into a generative model as auxiliary context to generate answers based on actual document support. In this embodiment, the RAG mechanism is optimized to support job-awareness and multi-level knowledge module retrieval, improving the accuracy and credibility of question-answering results.

[0027] Traditional intelligent question-answering systems and large models lack a deep understanding and dynamic perception of user role identities. As a result, when generating answers, the system is unable to consider the specific position of the questioner (such as risk internal control personnel, legal affairs specialists or retail banking salespersons), as well as the different access rights, understanding and expression methods of these positions for knowledge content. The relevant systems often provide a generalized answer that does not take into account role differences. This not only reduces the user experience, but may also lead to business errors because the answer exceeds the user's authority or does not meet their specific needs.

[0028] Moreover, although the current semantic retrieval system based on vector matching can understand the intention of the question to a certain extent, it seems powerless when faced with the complex and ever-changing financial knowledge system. For example, the relevant system usually stores all document fragments in a unified vector library for global similarity matching, ignoring the line attributes, scope of application and version control of the institutional documents. As a result, during the retrieval process, institutional content from different business lines that is outdated or not applicable to the current position may be mistakenly recalled and pieced together into the answer, resulting in problems such as "mixed knowledge modules" and "incorrect institutional references."

[0029] In addition, the financial industry has strict audit and traceability requirements for the output results of intelligent question-answering systems. However, the answers provided by most current question-answering systems lack a clear binding with the original institutional documents, and are unable to provide the specific institutional provisions, versions and source information on which the answers are based. This makes it impossible to effectively audit and verify the answers when disputes arise, and is not conducive to compliance inspections and responsibility tracing, which greatly limits the application of intelligent question-answering systems in the core business processes of financial institutions.

[0030] In order to solve the above technical problems, the embodiments of the present application provide corresponding solutions, which are described in detail below.

[0031] The embodiment of the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches provided in the embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 The hardware structure block diagram of a computer terminal for implementing a modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches is shown. Figure 1 As shown, the computer terminal 10 may include one or more processors (illustrated as 102a, 102b, ..., 102n in the figure) (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission module 106 for communication functions via a wired and / or wireless network connection. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art will understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0032] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry." The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuitry may be a single, independent processing module, or may be fully or partially integrated into any of the other components of the computer terminal 10. As discussed in the embodiments of the present application, the data processing circuitry functions as a processor control (e.g., the selection of a variable resistor terminal path connected to an interface).

[0033] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, implementing the above-mentioned modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include memory remotely located relative to the processor, and these remote memories can be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, corporate intranet, local area network, mobile communication network, and combinations thereof.

[0034] The transmission module 106 is configured to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the communications provider of the computer terminal 10. In one embodiment, the transmission module 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, the transmission module 106 may be a radio frequency (RF) module, which is configured to communicate with the Internet wirelessly.

[0035] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 .

[0036] It should be noted that, in some optional embodiments, the above Figure 1 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0037] In the above-mentioned operating environment, an embodiment of the present application provides an embodiment of a modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0038] Figure 2 This is a flow chart of a modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches according to an embodiment of the present application. Figure 2 As shown, the method includes the following steps:

[0039] Step S202 : obtaining a query vector of the target object, wherein the query vector includes a question vector for expressing the query requirement of the target object and a position vector for reflecting the position information of the target object in the financial branch.

[0040] In step S202 above, the query vector refers to converting the query requirements of the target object into a vector form that can be processed by a computer. In some embodiments of the present application, the query vector may be composed of a question vector and a position vector to guide the retrieval system and model to more accurately understand and respond to user queries:

[0041] (1) Question vector: a mathematical vector representation of the specific query requirements of the target object. For example, it can be generated by a semantic encoding model based on the natural language question input by the user. The role of the question vector is to provide semantic input for the retrieval stage to help the system understand the intention and context of the user's question. In some embodiments of the present application, when generating the question vector, the natural language question input by the user can be converted into a vector representation through a semantic encoding model. The semantic encoding model can be a deep learning model based on the Transformer architecture, such as BERT or its variants. In specific implementation, the question can be input into the encoder part of the model, and the vector output by the model will capture the key semantic features of the question. For example, if the user's query is "What is the latest policy for small and micro enterprise loans?", the semantic encoding model will generate a vector that not only contains the semantic information of loans and policies, but also implies the two key conditions of small and micro enterprises and the latest.

[0042] (2) Position vector: A mathematical vector that reflects the target object's position identity and authority in a financial branch. It can also be generated by a special model based on the user's position information. The role of the position vector is to limit the institutional boundaries of the retrieval and generation process to ensure that the answer content complies with the user's position authority and business scope. The construction of the position vector involves converting multi-dimensional information such as the target object's organizational attributes, position responsibilities and knowledge authority into a vector representation. In some embodiments of the present application, a vector space can be designed, in which each dimension represents a position feature, such as business field, department, responsibility level, etc., and a vector is defined for each position, in which the value of each dimension reflects the attribute of the position in the corresponding feature. For example, for a "risk compliance position", its vector has a high value in the "compliance" dimension and a high value in the "risk management" dimension, but may be zero or close to zero in the "product sales" dimension. This vector is used to filter and customize the response content during the retrieval and generation stages to ensure that the answer complies with the position authority scope and responsibility requirements.

[0043] When a user's query involves complex context, the question vector can be generated by leveraging contextual information such as the user's historical query history, operational behavior, and business scenarios to ensure it accurately reflects the user's true intent and contextual information, preventing irrelevant or semantically biased content from appearing in the search results. This can further enhance the question vector's semantic representation capabilities. For example, if the user is an account manager who frequently handles loan transactions, and the query is made during the approval process for a specific loan project, the system can convert this historical behavior and current scenario information into a vector and integrate it with the question vector to enhance the question vector's representation of the job context and business background.

[0044] Step S204 : determining the policy module accessible to the target object based on the position vector, wherein the policy module is a unit that logically divides policy documents according to the business areas of the financial branch.

[0045] In step S204, the policy module is a document set logically divided according to the financial branch's business areas, / or job responsibilities, and / or policy types. Each policy module focuses on a specific business area or job function and contains a set of relevant policy documents and operational guidelines. The purpose of the policy module is to provide a structured knowledge organization method, making it easier for the system and users to locate specific business rules and operational processes, thereby achieving accurate retrieval and effective utilization of knowledge.

[0046] In some embodiments of the present application, the system module can be determined by multi-dimensional job label mapping. Specifically, when determining the system module accessible to the target object, the system first parses the multi-dimensional job label information carried in the job vector, including but not limited to job title, department, business line, knowledge authority level, etc., and then maps this information to a predefined system module access control list, which records in detail the access rights of each position to different system modules. For example, for the "risk compliance" label in the job vector, the system will query the control list for system modules related to risk compliance, such as "credit risk management", "fund flow monitoring policy", etc., so as to determine the scope of content accessible to the target user.

[0047] In addition to using job tags for direct permission mapping, this application can also introduce a system module permission management mechanism based on the knowledge graph. The knowledge graph can not only express the logical relationship between system documents, but also embed job access rules and the scope of application of the system. Specifically, the system locates the role node of the user (i.e., the target object) in the graph based on the job vector, and identifies the accessible path under the node through the graph traversal algorithm, that is, the set of system modules that the user can query. For example, the "risk compliance position" node may have access to specific modules under the graph such as "internal control process" and "compliance report". The system automatically identifies these paths and limits the search scope to ensure that the answer content is aligned with the user's permissions.

[0048] The above steps solve the problem that existing question-and-answer systems are unable to provide customized answers for users with different positions and permissions. By combining the position vector with the system module access rules, the system can limit the scope of retrieval and answer generation based on the user's specific position and knowledge permissions, avoiding cross-departmental information leakage or access to unauthorized system content, and ensuring the compliance and security of financial knowledge services.

[0049] When a user has multiple identities or cross-departmental responsibilities, in order to determine their comprehensive access rights and avoid conflicts or omissions in permissions, the following steps can also be performed: the system maintains a cross-position permission system, which records the permission inheritance and merging rules between different positions. For example, a "Risk Compliance Officer" may have access rights to both the "Credit" and "Fund Flow Monitoring" modules, while a "Senior Risk Compliance Officer" may inherit the former's permissions and also have additional access rights to the "Cross-border Risk Assessment" module; when processing such position vectors, the system will automatically merge all relevant permissions of the user according to the permission system to form a comprehensive access permission list.

[0050] It's important to note that when a user has complex or changing roles, the system can employ a dynamic permission resolution algorithm to calculate in real time the set of institutional modules they can access. Specifically, based on the identity information in the position vector, this algorithm leverages the graph structure of a knowledge graph embedded with position access rules and institutional scopes to automatically analyze permission inheritance paths and cross-permission scopes, ensuring that access decisions for each module are based on the user's entire current set of roles and permissions. For example, if a user is both a "Retail Banking Manager" and an "Internal Control and Compliance Advisor," the system will resolve the permissions for both roles separately and then merge them to determine a comprehensive list of institutional module access permissions.

[0051] When a user's identity changes dynamically or they operate across departments, the system also features role-context awareness, automatically identifying the user's current task scenario and temporary role and dynamically adjusting their access permissions. For example, when a "Retail Banking Account Manager" performs a specific "Fund Flow Monitoring Investigation" task, the system can temporarily enhance their access permissions to the "Fund Flow Monitoring" module, ensuring they have access to all the regulatory information necessary to perform the task.

[0052] In some embodiments of the present application, the system modules accessible to the target object can be determined by the following steps: obtaining multidimensional labels corresponding to all system modules, wherein the multidimensional labels include business domain labels for limiting the scope of application of system documents, line attribute labels for refining business lines under the business domain, and position applicability labels for restricting access rights; determining the similarity between the position vector and the multidimensional label, and comparing the position authority information in the position vector with the position applicability label to obtain a comparison result; determining the system modules accessible to the target object based on the similarity and the comparison result, wherein each system module corresponds to a sub-index library, and the sub-index library is used to store the system documents corresponding to the system module.

[0053] Specifically, multi-dimensional tags can include business domain tags, line attribute tags, and position applicability tags, which are used to accurately describe the scope of application of the policy document, subdivide the business domain, and access rights.

[0054] (1) Business area label: used to identify the main business scope covered by the system document, such as "retail banking", "corporate banking", "compliance supervision", "risk management", etc. This label helps the system understand the macro business area of the document.

[0055] (2) Line attribute labels: further refine the business area labels, involving specific business lines or operating procedures, such as "personal loans", "credit cards", "fund flow monitoring", "credit review", etc., which are used to distinguish the meso-business attributes of institutional documents.

[0056] Line attribute tags can be determined based on business domain tags, combined with the specific content of policy documents, business processes, and organizational structure. This is explained below with some specific examples. Assume there is a collection of policy documents for the business domain "Retail Banking." You want to further refine the document classification within this domain to support more accurate retrieval and permission management. The specific steps are as follows:

[0057] 1) Analyze the content of regulatory documents: Perform semantic analysis and text mining on all regulatory documents in the "retail banking" field to identify key processes, products, and services that frequently appear in the documents. For example, frequently mentioned business lines such as "personal loan approval," "credit card issuance," and "savings account management" can be identified from "retail banking" regulatory documents.

[0058] 2) Develop line attribute labels: Based on the identification results, a set of line attribute labels are developed for the "retail banking" field, including but not limited to "personal loans," "credit cards," "savings accounts," and "wealth management." Each label reflects a specific business line or operational process in the "retail banking" field, providing detailed standards for subsequent permission control and document retrieval.

[0059] 3) Label documents with line attribute tags: Each institutional document in the "retail banking" field will be labeled with the corresponding line attribute tag based on the actual business line it covers. For example, an institutional document on the "personal loan approval process" will be labeled with the "personal loan" line attribute tag, while another document on "credit card fraud detection" will be labeled with both "credit card" and "anti-fraud" tags.

[0060] 4) Maintain the mapping relationship between line attributes and business areas: Maintain a mapping table or relationship network in the system to record the association between each line attribute label and the label of the corresponding business area. For example, lines such as "personal loans," "credit cards," and "savings accounts" are all mapped to the business area "retail banking," forming a clearly defined labeling system.

[0061] 5) Build an index and retrieval algorithm: Policy documents with line attribute tags are stored in separate sub-indexes, each corresponding to a specific business line. During actual retrieval, the system quickly locates the corresponding broad category index based on the business domain tags in the user's query. It then further filters the document collection based on the line attribute tags, achieving precise targeting and efficient recall.

[0062] (3) Position applicability label: defines the access rights and applicable objects of the document, such as "all employees", "specific positions", "internal audit", "risk control department", etc., to ensure that only users with corresponding permissions can access the corresponding documents, maintaining the security and compliance of information.

[0063] In some embodiments of the present application, the system first obtains the multidimensional labels of all institutional modules and converts them into vector representations. It then calculates the similarity between the position vector and the multidimensional label vector of each institutional module (e.g., using cosine similarity), and uses this similarity score to determine the institutional modules accessible to the user. Through vectorization and similarity calculation, the system can quickly and accurately determine the institutional modules accessible to a user based on their specific position and permissions, avoiding the inefficiency and mismatch risks associated with traditional keyword- or rule-based permission management and improving the precision and flexibility of permission control.

[0064] When determining the system modules accessible to the target object, the system can also compare the position authority information in the position vector with the position applicability label of the system module. If the comparison result shows that there is a matching relationship between the user's position and the applicability label of the system module, and the similarity calculation result is higher than the preset authority threshold, the system determines that the user can access the corresponding system module.

[0065] In order to facilitate understanding of the above process, some specific embodiments are explained below.

[0066] In a large commercial bank, assume that user A is a senior manager in the credit department. His / her work scope covers multiple business areas, including but not limited to corporate loans, personal consumption loans, and risk management. User A needs to query specific details about the corporate loan approval process.

[0067] Step 1: Get multi-dimensional labels.

[0068] The system first obtains multi-dimensional tag information for all system modules within the bank, including but not limited to business area tags, line attribute tags, and position applicability tags. For example, the multi-dimensional tags for the "Corporate Loans" system module might include: "Corporate Banking" (business area tag), "Corporate Loan Approval" (line attribute tag), and "Credit Manager and Above" (position applicability tag). The multi-dimensional tags for the "Risk Management" system module might include: "Risk Management" (business area tag), "Loan Risk Assessment" (line attribute tag), and "Entire Risk Management Department" (position applicability tag).

[0069] Step 2: Calculate similarity and compare permissions.

[0070] Calculating similarity: The system calculates the similarity between user A's job vector and the multidimensional label vector of each system module. For "Senior Manager of the Credit Department," the job vector has higher weights related to "corporate loans" and "risk management." The system compares the job vector with the business domain and line attribute label vectors of each system module to obtain a series of similarity scores.

[0071] Comparing permissions: The system compares User A's position permissions (e.g., "Senior Manager, Credit Department") with the position applicability tags in the policy module to check for a match. For the "Corporate Loans" and "Risk Management" policy modules, since "Senior Manager, Credit Department" falls within the applicability scopes of "Credit Department Manager and Above" and "Entire Risk Management Department," the comparison result is a match.

[0072] Step 3: Identify accessible institutional modules.

[0073] Based on the similarity scores and permission comparison results from step 2, the system determines the policy modules that User A can access. In this example, the "Enterprise Loans" and "Risk Management" policy modules have high similarity scores, and their job applicability tags match User A's job permission information. Therefore, the system identifies these two policy modules as accessible to User A.

[0074] Step 4: Index positioning and document retrieval.

[0075] The system identifies the target system modules and locates their corresponding sub-indexes: the "Enterprise Loans" system module's corresponding sub-index, and the "Risk Management" system module's corresponding sub-index. Within these sub-indexes, the system can use semantic search technology, based on the vector representation of User A's query, to retrieve system documents related to the "Enterprise Loan Approval Process." Because the sub-indexes have already been pre-screened by business line and authority, the search results are more accurate and secure.

[0076] Step S206: determining a subgraph region corresponding to the question vector in the knowledge graph, wherein the knowledge graph is used to represent the logical structure between institutional documents.

[0077] In the above step S206, the knowledge graph is a data structure used to represent the logical relationships and structural dependencies between financial system documents, such as entity relationships, system references, business process chains, etc. Its role is to provide a structured view of the system content, helping the system understand the internal connections of the system text, so as to make more reasonable and accurate decisions in the retrieval and generation stages.

[0078] A subgraph area is a local structure or set of nodes in the knowledge graph that is associated with a specific question vector. It is used to limit the focus of the intelligent question-answering system in the process of institutional retrieval and ensure that the generated answer content conforms to the semantic scope and specific context of the question. In some embodiments of the present application, when a user asks a question, the system first maps the question vector to one or more semantic anchors in the knowledge graph. These semantic anchors can be the nodes in the graph that are most relevant to the topic of the question. For example, "personal loan approval process" may point to the "personal loan" or "approval" node. Subsequently, the system can perform a multi-hop path search in the graph to explore the set of nodes reached from the semantic anchor point through logical edges such as "reference", "related", and "inheritance" to form a subgraph area that matches the question vector.

[0079] In the complex institutional structure of financial branches, the above search strategy can help the system understand the deep semantics of the questions, avoid false recalls caused by surface semantic similarities, ensure that the generated answers are highly relevant to user needs, and follow the logical relationship between institutional documents, thereby improving the accuracy and compliance of questions and answers.

[0080] It should be noted that the knowledge graph can also be logically partitioned according to business areas or job permissions. Each partition contains a set of nodes and edges directly related to the area or permission. After determining the business area or position of the user's query, the system only performs multi-hop path search in the corresponding partition, rather than in the entire graph, further reducing the computational complexity of the search.

[0081] When determining subgraph regions, the system also adjusts the weights of nodes and edges in the graph based on the urgency and relevance of the user's query. For example, if the question is "What are the latest requirements regarding the policy on monitoring capital flows?", nodes and edges directly related to "capital flow monitoring" (such as "Policy Announcements" and "Latest Revisions") will be given higher weights, while indirectly related or outdated nodes will be given lower weights. The system then sorts multi-hop paths based on their weights, prioritizing subgraph regions formed by paths with higher weights, thereby ensuring the timeliness and authority of the answer.

[0082] When building a multi-hop path search and weight adjustment mechanism for a knowledge graph, determining the direct and indirect relevance of nodes and edges is key to ensuring the timeliness and authority of the answer content. In some embodiments of this application, the weights of individual nodes and edges can be adjusted based on the urgency and relevance of the user's query:

[0083] Step 1: Construct a Term Frequency-Inverse Document Frequency (TF-IDF) weighting model. First, perform text analysis on all nodes in the knowledge graph. For example, based on the topic "Fund Flow Monitoring Policy" (i.e., the query topic extracted from the question vector), calculate the TF-IDF value of each node. Directly related nodes (such as "Policy Announcement" and "Latest Revision") tend to have a high frequency of mentions of the topic in documents and are relatively unique in the graph, resulting in relatively high TF-IDF values. Indirectly related nodes (such as "Customer Due Diligence"), while also related to the topic, have lower frequency of mentions and lower uniqueness, resulting in smaller TF-IDF values.

[0084] Step 2: Define a time weight function. For example, for the keyword "latest requirements," the system needs to identify which nodes contain the most recently updated information. It should be noted that if the question vector does not contain time-related keywords, the time the question vector was raised can be used as a benchmark. Specifically, a time weight function can be defined that takes the node's update time or version release date as input and outputs a weight value that quantifies the correlation between the node's time and the time indicated by the question vector. In the above example, the newer the information, the higher the weight.

[0085] Step 3: Implement graph traversal and weight accumulation algorithm.

[0086] a. Determine the starting point: Map the user question vector to the knowledge graph and find the initial node most relevant to the question. In the above example, this could be the "Fund Flow Monitoring Policy" node.

[0087] b. Weight initialization: Assign preliminary weights to the starting node and all directly connected edges. Directly related edges (such as those pointing to "Policy Announcement" and "Latest Revision") will receive higher initial weights, while indirectly related edges will receive lower weights.

[0088] c. Multi-hop path weight update: Starting from the starting node, a multi-hop traversal is performed along all edges. During the traversal, the weight of each node and edge is dynamically adjusted based on its relevance to the query topic (TF-IDF value) and timeliness (output value of the time weight function). For example, the weight of a node is the product of the average weight of all edges reaching it and the TF-IDF value of the node. The weight of an edge is based on the weighted average of the TF-IDF values of its connected nodes and the result of the time weight function.

[0089] d. Weight propagation and reduction: During graph traversal, weights are gradually decayed as the number of hops increases, reducing the weights of indirectly related nodes and ensuring that the system's focus remains on the nodes most directly relevant to the query. For example, an exponential decay function can be used, reducing the weight by 50% with each hop.

[0090] e. Subgraph area determination: After the traversal is completed, the system will sort all reachable nodes according to the comprehensive weight of the nodes (initial weight + TF-IDF value + time weight) and select the top N nodes with the highest weights to form the subgraph area. The size of N can be dynamically adjusted based on the system design and query complexity.

[0091] Step 4: Prioritization and Answer Generation: The system prioritizes nodes with higher weights based on the weights of the nodes in the subgraph region. This ensures that answers are based on the latest and most relevant information. Furthermore, higher-weighted nodes serve as the primary source of information, while indirectly related or outdated nodes are used only as supplementary information or background knowledge, enhancing the completeness and depth of the answer.

[0092] In some embodiments of the present application, a knowledge graph may be constructed in the following manner: extracting entities from a collection of institutional documents, wherein the entities include institutional clauses; determining the relationships between the entities, and determining the edges of the knowledge graph based on the relationships, wherein the relationships include one of the following: reference relationships, applicability relationships, inheritance relationships, and substitution relationships between institutional clauses; and constructing a knowledge graph based on the entities and edges.

[0093] Entities refer to key components extracted from various institutional documents, including but not limited to institutional clauses, policies and regulations, process steps, job roles, etc. Relationships, namely the relationships between entities, such as "reference," "apply," "inherit," and "replace," are key to building the logical structure of the knowledge graph. Their role is to connect entities, forming edges in the knowledge graph, reflecting the logical dependencies and evolutionary trajectories between institutional clauses, and helping the system understand the dynamic changes and semantic connections of the system. They include:

[0094] (1) Reference relationship: refers to a system clause directly citing another clause as its basis or reference, for example, "loan approval process" citing "customer due diligence requirements" as a necessary step before approval.

[0095] (2) Applicability: Indicates that an entity or clause is applicable to a specific position or business scenario, for example, “loan approval process” is applicable to “credit department manager”.

[0096] (3) Inheritance relationship: describes the inheritance relationship between the new and old system clauses, which means that the new clauses continue or modify the content of the old clauses to some extent, for example, the "2024 version of the loan approval process" inherits from the "2023 version of the loan approval process".

[0097] (4) Substitution relationship: refers to a new clause replacing an old or outdated clause and becoming the new implementation standard, for example, the “Newly Revised Funds Flow Monitoring Policy” replaces the “2022 Edition of the Funds Flow Monitoring Policy”.

[0098] After constructing the basic knowledge graph, which includes the reference, application, inheritance, and substitution relationships between institutional clause entities, we can further integrate the position information and authority information of financial branches. Specifically:

[0099] (1) Introducing position nodes: Add “positions” as new entity nodes in the knowledge graph, including but not limited to “teller”, “account manager”, “risk compliance specialist”, etc. These position nodes will be listed alongside other entity nodes such as system clauses and process nodes, forming part of the graph to represent specific business operators and information consumers.

[0100] (2) Define the applicable relationship: Create an edge of the "applicable relationship" between the position node and the system clause. This relationship indicates which positions a certain system or process is applicable to. For example, establishing an applicable relationship between "loan approval process" and "credit manager" indicates that "loan approval process" is specifically designed for the position of "credit manager". In this way, the system can ensure that only the system clauses that match the position's permissions are retrieved and used.

[0101] (3) Permission labels and edge weights: Permission labels are added to each position node and system clause node. These labels can include "readable", "editable", "requires review", etc. At the same time, permission weights are added to the edges of applicable relationships to indicate whether different positions can access or apply the system clause. Permission weights can be binary (accessible / inaccessible) or multi-level (such as read-only, editable, full control, etc.).

[0102] (4) Constructing a permission filtering mechanism: During the graph retrieval and answer generation process, the system needs to construct a permission filtering mechanism. First, based on the user's identity information, the position node and the corresponding permission label are determined; when searching for institutional clauses that match the question vector, the system only allows access to clause nodes that have a direct or indirect applicable relationship with the user's position node and that are permitted by the permission weight, to ensure that the generated answer content complies with the user's access rights and job responsibilities.

[0103] (5) Dynamic permission adjustment: Considering that job permissions may be adjusted over time, over projects, or over policy changes, the knowledge graph also has a dynamic permission adjustment mechanism. For example, the update of job permissions is monitored through event listeners. Once it is detected that the permissions of the "Credit Manager" position have changed (such as the addition of review rights), the applicable edge weights involving the position in the graph are automatically updated to ensure that the system can reflect permission changes in a timely manner and avoid outdated permission control affecting the accuracy of questions and answers.

[0104] In some embodiments of the present application, a subgraph area corresponding to a question vector can be determined in a knowledge graph in the following manner: matching the question vector with the knowledge graph to obtain a first graph node, wherein the first graph node includes a node whose similarity with the semantic features of the question vector is greater than or equal to a first preset threshold; searching for a graph path starting from the first graph node based on the edges of the knowledge graph, and stopping the search when a stopping condition is met to obtain a second graph node set, wherein the stopping condition includes that the similarity between the semantic features of the second graph node and the question vector is less than a second preset threshold, and the second preset threshold is less than the first preset threshold; determining a subgraph area based on the first graph node and the second graph set.

[0105] The first graph node is a node in the knowledge graph whose semantic feature similarity with the question vector is greater than or equal to a first preset threshold. It is considered a candidate node highly associated with the question and is used to provide a starting point for the graph path search. The second graph node set is the set of all nodes that meet the stopping condition when performing a graph path search starting from the first graph node. It is used to limit the subgraph range with a high degree of relevance to the user's question to ensure that the search results do not deviate from the semantic core of the user's query. The subgraph area is a local area of the knowledge graph related to the user's question, determined based on the first graph node and the second graph node set.

[0106] Specifically, the system analyzes the first graph node, identifies the edges it connects to and the nodes it points to, and then performs multi-hop expansion of the graph path based on edge types (such as reference, applicability, inheritance, and substitution relationships) to explore nodes that are directly or indirectly associated with the semantic features of the question. During the graph path expansion process, the system sets a stopping condition, such as stopping the search when the similarity between the second graph node and the question vector falls below a second preset threshold. The second graph node set represents the search boundary.

[0107] In the institutional knowledge graph of financial branches, the edge type not only reflects the complex logical relationships between entities, but also implies different strategies for retrieval and knowledge expansion. In some embodiments of this application, different expansion strategies can be adopted based on the edge type (reference, application, inheritance, substitution relationship) to ensure the semantic coherence and structural integrity of the graph path, including:

[0108] (1) Reference relationship edge expansion strategy: When encountering a reference relationship edge, the system should prioritize deep searching along this edge to ensure that the referenced entity or institutional clause is found. Reference relationships usually indicate direct dependencies between institutions, for example, "loan approval process" references "customer due diligence requirements." During the expansion process, the system will mark this path as a "direct reference path" and assign a higher similarity weight to subsequent nodes found along this edge because they are directly related to the institutional clause in the query and provide direct information required to answer the question.

[0109] (2) Applicable relationship edge expansion strategy: Applicable relationship edges point to other entities related to the current entity. For example, "customer due diligence requirements" may be applicable to the "retail banking" and "corporate banking" departments. When the system encounters such edges, it will adopt a breadth-first search strategy (BFS), that is, first explore all directly related entities, and then go deeper layer by layer. The breadth exploration of applicable relationships helps to build a more comprehensive semantic background and provide contextual information that users may need. However, the expansion of applicable relationship edges should be limited in depth to avoid excessively distant entities from interfering with the results, ensuring the directness and accuracy of the answers.

[0110] (3) Inheritance Edge Expansion Strategy: Inheritance edges indicate that a new entity (or a new version of a system) inherits the content and attributes of an old entity. For example, the "2024 version of the loan approval process" inherits from the "2023 version of the loan approval process." When the system encounters an inheritance edge, it adopts a time-first strategy, prioritizing exploration and consideration of entities that are closer in time (i.e., newer versions). If no relevant content is found in the newer version, the system will go back to the old entity and continue searching. In addition, the system can also record the inheritance path so that the trajectory of system evolution can be reflected in the final answer, enhancing the authority and traceability of the answer.

[0111] (4) Replacement relationship edge expansion strategy: A replacement relationship edge indicates that one entity (or institutional clause) completely or partially replaces another entity. For example, the "New Fund Flow Monitoring Policy" replaces the "2022 Edition of the Fund Flow Monitoring Policy". When encountering a replacement relationship edge, the system should adopt a newest priority strategy and jump directly to the replaced entity unless the user explicitly asks for information about the old entity. The processing of replacement relationships requires special attention to version control and timestamps to ensure that the system does not return outdated content. At the same time, the answer should clearly mark whether the entity based on the replacement relationship is the new entity or the old entity in the replacement relationship to meet compliance and auditability requirements.

[0112] Based on the expansion strategies of the above four edge types, the system can use the following methods in combination when searching for multi-hop paths:

[0113] (1) Priority sorting: The system can set different search priorities for each edge type. For example, the reference relationship edge has the highest priority because it is directly related to the core content of the problem; the second is the inheritance relationship edge and the substitution relationship edge because they show the evolution and update of the system; the applicability relationship edge has a relatively low priority and is used to construct the semantic background.

[0114] (2) Dynamic Adjustment of Depth and Breadth: In some embodiments of the present application, different search strategies can be set for different types of edges. For example, for reference and substitution relationship edges, the system can adopt a depth-first strategy to deeply explore directly related entities; for applicability and inheritance relationship edges, a breadth-first strategy can be adopted to first explore all directly related entities before going deeper. It should be noted that the system can dynamically adjust the search depth according to the edge type (i.e., determine the stopping condition of the search strategy based on the edge type) to avoid redundant searches and information overload.

[0115] (3) Semantic similarity and version timeliness weights: During the expansion process, for each type of edge, the system can combine semantic similarity and version timeliness (such as release time, validity period) to calculate the weight to decide whether to continue searching along the edge. For example, for an outdated replacement relationship edge, even if the semantic similarity is high, its weight should not exceed the weight of the new version edge.

[0116] Step S208: Retrieve the document paragraph corresponding to the sub-graph area from the sub-index library corresponding to the system module, and input the document paragraph into the large model to generate a target answer corresponding to the query vector.

[0117] In the above step S208, the document paragraph is the smallest retrievable unit after the institutional document is segmented. Each paragraph is converted into a vector representation and stored in the corresponding sub-index library. Its function is to provide the minimum granularity information retrieval capability to ensure the accuracy of the retrieval results and the readability of the recalled documents.

[0118] The large model is responsible for understanding and parsing the question vector entered by the user, and simultaneously combining it with document paragraphs retrieved from the knowledge graph to generate the answer that best meets the query requirements and matches the job information. When generating answers, the large model can perceive and understand the context of the user's role and determine the style, depth, and applicability of the answer based on the job vector. This ensures that the generated answer meets both the professional standards of the financial industry and the specific requirements and knowledge requirements of the position, thereby improving the quality of the answer and user satisfaction.

[0119] In some embodiments of the present application, accurate target answers can be generated by combining a large model with job information. Specifically:

[0120] (1) Obtaining a job vector: The system first generates a job vector based on the user's identity and job information. This vector contains the user's authority boundaries, business context, and job characteristics, and is used to guide the retrieval and generation process.

[0121] (2) Determine the accessible system modules: By comparing the job vector with the vectors of each system module, the system determines the system modules that the user can access, ensuring that the document paragraphs within the search range match the user's job permissions, and preventing improper exposure of unauthorized information.

[0122] (3) Document retrieval based on subgraph regions: Locate the subgraph region corresponding to the question vector in the knowledge graph. This subgraph region represents the specific knowledge content and logical structure required to answer the question. The system retrieves the most relevant document paragraphs in the sub-index library of the institutional module related to this region, while ensuring that the metadata tags of these paragraphs (such as version, regional applicability, etc.) match the user's needs.

[0123] (4) Constructing control instructions and constraints: The system constructs control instructions and constraints based on the path information of the retrieved document paragraphs and the user's position information. The control instructions guide the model to generate answers in the logical order of the knowledge graph, while the constraints ensure that the generated answers only contain information that the user has access to.

[0124] For example, if a credit manager asks about the "loan approval process", the system's control instructions may be: "Generate detailed steps on the loan approval process, with particular emphasis on risk assessment and guarantee conditions", and set restrictions: "Only reference documents applicable to the credit manager position, and the documents must be within the validity period."

[0125] (5) Input fusion and answer generation: The user question vector, control instructions and constraints are input into the big model, and the model generates the target answer based on this information. In this process, the big model not only processes natural language questions, but also integrates structured knowledge, job context and authority control to ensure the professionalism, accuracy and compliance of the answer content.

[0126] In some embodiments of the present application, based on the subgraph area constructed by the user question, the system needs to identify the institutional module to which the nodes in the subgraph area belong. For example, each node is labeled with a business line or institutional category in the knowledge graph, and the system determines the institutional module to which the node belongs through a label matching mechanism. Specifically, the node labels in the subgraph area are compared with the labels of each sub-index library to determine which sub-index libraries contain document paragraphs related to the subgraph area. For example, if the subgraph area contains the "loan approval" and "risk compliance" nodes, the system will locate the "credit management sub-index library" and the "risk internal control sub-index library".

[0127] In an embodiment of the present application, the system modules within the scope of the user's access rights are determined based on the user's job identity, ensuring that subsequent retrieval operations are only performed in the knowledge fields that the user has access to, and based on the user's question vector, the retrieval scope is further refined in the sub-index library of the determined system module, and the logical boundaries of the retrieval are limited by matching the sub-graph areas in the knowledge graph, ensuring that the recalled document paragraphs are highly relevant to the specific semantic features of the query. Through the dual screening of job permissions and question semantics, the system can accurately determine which system documents can be accessed by users in specific positions and which content is most relevant to the problem, greatly enhancing the security and compliance of the system. The first screening limits the physical scope of the retrieval through the system module, while the second screening further refines the logical boundaries of the retrieval, significantly reducing the time and computing cost of the retrieval operation, and improving the system response speed and overall performance.

[0128] The filtered document paragraphs are input into the pre-trained large language model. Combined with the vector representation of the user question, a target answer corresponding to the query vector can be generated. Specifically: the guidance information of the large model is obtained, wherein the guidance information includes the path information of the document paragraph in the knowledge graph and the position information of the target object. The path information is used to indicate the institutional logic and structural order followed by the large model when generating the answer; the target template is constructed based on the guidance information, wherein the target template includes the control instructions corresponding to the path information and the restriction conditions corresponding to the position information; the control instructions and restriction conditions are processed by the large model to obtain the target answer.

[0129] Guidance information is a data package containing the path information of the document paragraph in the knowledge graph and the target object's position information. Its role is to guide the large model to follow specific institutional logic and structural order when generating answers, while taking into account the user's role and permissions to ensure that the answers are both accurate and legal.

[0130] (1) Path information: The specific path from the question anchor point (i.e., the first graph node) to the recalled document paragraph in the knowledge graph, which is used to control the logical process and structural framework of the large model to generate answers.

[0131] (2) Job information: The job context of the user’s question, which is used to consider job-specific terminology, concerns, and authority scope when generating answers, ensuring the job-appropriateness of the answer content.

[0132] The target template is constructed based on the guidance information and contains control instructions and constraints. It is used to format input during large-scale model generation, ensuring that the generated answers follow a specific logical structure and semantic boundaries while also satisfying job authority constraints. Control instructions refer to instructions in the target template that instruct the large-scale model to generate answers according to the institutional logic and structural sequence defined in the path information. Constraints refer to boundary conditions in the target template that limit the content of answers generated by the large-scale model. These are primarily based on the user's job information and ensure that the answers meet not only semantic requirements but also job authority and business specifications.

[0133] In some embodiments of the present application, the complete path information of a document paragraph in the knowledge graph can be extracted based on the determined subgraph region, including the starting node, intermediate reference nodes, end node, and edge types along the path (such as reference, application, inheritance, etc.), to form a path description string. The user's position information (such as "Credit Manager") is converted into a position vector and combined with the path description string to form a complete guidance information data package. For example, if a "Credit Manager" asks about the "loan approval process" under the 2024 version of the institutional framework, the guidance information may include the path information: "Loan Approval Process → Customer Due Diligence Requirements → Approval Authority Settings" and the position information: "Credit Manager."

[0134] The system converts path information into control instructions. These instructions can be specific prompt words, sentence frameworks, or model control parameters, which guide the large model to generate answers based on the knowledge structure. For example, for the path information "Loan Approval Process → Customer Due Diligence Requirements → Approval Authority Settings," the control instruction might be: "When describing the loan approval process, please include the specific steps for customer due diligence requirements and approval authority settings."

[0135] Furthermore, role-based restrictions must be set in the target template. These restrictions inform the big model user of the role's scope of authority and terminology, ensuring that the answers are appropriate for the role. For example, a restriction might be set as follows: "Answers must not contain content related to internal audit processes, and only credit department managers can access policy clauses."

[0136] By constructing target templates, the system can control the content boundaries and logical processes of large models during the generation process, solve the problems of lack of structural constraints and job adaptability of answer content, and at the same time meet the financial industry's compliance and accuracy requirements for knowledge dissemination.

[0137] After retrieving the document paragraph corresponding to the sub-graph area in the sub-index library corresponding to the system module, the following steps can also be performed: obtaining the metadata tag corresponding to the sub-node of the sub-graph area, wherein the metadata tag includes the effective time, abolition time and applicable region of the system document; when there are multiple versions of the system document corresponding to the sub-node, determining the matching score between the system document corresponding to each version and the query vector based on the metadata tag, wherein the matching score is used to quantify the applicability of the document paragraph corresponding to each version; and determining the target document paragraph for generating the target answer based on the matching score.

[0138] The matching score between the institutional document corresponding to each version and the query vector is determined based on the metadata tag. Specifically: the query time corresponding to the query vector is obtained, and a first indicator is determined based on the query time, the effective time and the abolition time corresponding to the version, wherein the first indicator is used to quantify the timeliness of the version; the query area corresponding to the query vector is obtained, and a second indicator is determined based on the query area and the applicable area of the version, wherein the second indicator is used to quantify the regional compliance of the version; a third indicator between the question vector in the query vector and the institutional document of the version is determined, wherein the third indicator includes semantic similarity; and the matching score is determined based on the first indicator, the second indicator and the third indicator.

[0139] Metadata tags refer to additional information added to nodes in the knowledge graph (especially those for policy documents) to describe the node's characteristics, such as the policy document's effective date, expiration date, and applicable regions. When multiple versions of a policy document exist, the matching score is calculated by the system to measure the correlation between each version's document paragraph and the user's query vector, thereby quantifying the applicability of each version.

[0140] Specifically, (1) during the institutional content processing stage, complete version metadata tags can be added to each institutional clause and document paragraph, including but not limited to: version number (Version ID), effective time (Start Time), abolition time (End Time), applicable region (Region), version applicable rules (Version Policy), etc. These version tags are bound to the nodes in the institutional map to form a time-series "institutional evolution path map" that can support cross-version structure alignment, inheritance relationship traceability and other functions.

[0141] (2) During the question-and-answer process, the system semantically analyzes the user's question content and contextual information, and automatically identifies the time information, position information and geographical background implied in the question. For example, if a user raises a question under the identity of "customer manager of Beijing branch in October 2024", the system will automatically limit the question to the document subset that should be adapted to "the system version applicable to the Beijing region, which is in effect in October 2024" based on the user's organizational structure and job responsibilities.

[0142] Specifically, the system can score candidate policy versions (i.e., matching scores) using the following version matching formula. This is achieved by calculating the semantic similarity between each version of the policy document and the user's question vector, and combining it with the applicability information about time and region in the metadata tags to derive the matching score for each version:

[0143]

[0144] in, It's time for questions. is the matching score, The target object's tissue area, The version applicable area of the system document. Indicates the degree of match between the semantics of the question vector and the semantics of the document paragraph; 、 、 is the weight parameter.

[0145] The above formula includes factors such as semantic similarity, time window matching, and regional applicability. The system ultimately selects the version with the highest score to generate content.

[0146] (3) To ensure that the generated results have audit capabilities and temporal context consistency, an "explicit version binding generation template" can also be designed, that is, a structured prompt is automatically appended to the end of the answer content generated by the large model, such as: This answer is generated in accordance with Article 7 of the "Guidelines for Credit Customer Classification" (Revised Edition, September 2024), applicable to regional account managers in City A, and the current version is effective from September 1, 2024 to August 31, 2025. This mechanism not only enhances the credibility and authority of the answer content, but also provides a basis for systematic audit backtracking and responsibility tracking. Even if the system content is updated again in the future, the historical question and answer records can still be restored and compared based on the version annotation, supporting accountability management and operation review during the system transition period.

[0147] Through the above steps S202 to S208, the position information and query requirements are integrated, and a query vector containing a question vector and a position vector is generated to screen out the system module that matches the target object position, and locate the subgraph area corresponding to the question vector in the knowledge graph. Then, in the sub-index library corresponding to the selected system module, a refined document paragraph retrieval is performed, and these paragraphs are input as context into the large model to generate answers, thereby achieving the purpose of accurately matching user queries with accessible system modules and accurately defining the logical boundaries of semantic retrieval, thereby achieving the technical effect of generating enhanced position adaptability and system compliance of question and answer content, improving retrieval efficiency and answer accuracy, and thus solving the technical problem that the question and answer system adopted by the related technology lacks recognition of user position identity and cannot provide matching answers according to the different position responsibilities and authorities of financial branches.

[0148] The present application also provides a modular knowledge graph and retrieval-enhanced big model fusion interaction system for financial branches, including an interactive device and a server, wherein the interactive device is connected to the server and is used to obtain a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirements of the target object and a position vector for reflecting the position information of the target object in the financial branch; the server is connected to the interactive device and is used to determine the institutional module accessible to the target object based on the position vector, wherein the institutional module is a unit for logically dividing institutional documents according to the business areas of the financial branch; a subgraph area corresponding to the question vector is determined in the knowledge graph, wherein the knowledge graph is used to represent the logical structure between institutional documents; a document paragraph corresponding to the subgraph area is retrieved in the sub-index library corresponding to the institutional module, and the document paragraph is input into the big model to generate a target answer corresponding to the query vector.

[0149] It should be noted that the modular knowledge graph and retrieval-enhanced large model fusion interactive system for financial branches is used to execute Figure 2 The modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches is shown. Figure 2 The relevant explanations in the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches also apply to the above-mentioned modular knowledge graph and retrieval-enhanced large model fusion interaction system for financial branches, and will not be repeated here.

[0150] Figure 3 This is a structural diagram of a modular knowledge graph and retrieval-enhanced large model fusion interaction device for financial branches according to an embodiment of the present application, such as Figure 3 As shown, the device includes:

[0151] An acquisition module 302 is configured to acquire a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirement of the target object and a position vector for reflecting the position information of the target object in a financial branch;

[0152] A determination module 304 is configured to determine, based on the position vector, a policy module accessible to the target object, wherein the policy module is a unit that logically divides policy documents according to the business areas of the financial branch;

[0153] A matching module 306 is configured to determine a subgraph region corresponding to the question vector in a knowledge graph, wherein the knowledge graph is used to represent the logical structure between policy documents;

[0154] The interaction module 308 is used to retrieve the document paragraph corresponding to the sub-graph area in the sub-index library corresponding to the system module, and input the document paragraph into the large model to generate a target answer corresponding to the query vector.

[0155] It should be noted that Figure 3 The modular knowledge graph and retrieval enhanced large model fusion interactive device for financial branches is used to perform Figure 2 The modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches is shown. Figure 2 The explanations in the interactive method of modular knowledge graph and retrieval-enhanced large model fusion for financial branches also apply to Figure 3 The modular knowledge graph and retrieval-enhanced large model fusion interactive device for financial branches shown will not be described in detail here.

[0156] An embodiment of the present application also provides an electronic device, which includes a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the steps of the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches in each embodiment of the present application.

[0157] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the steps of the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches in each embodiment of the present application by running the computer program.

[0158] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches in each embodiment of the present application.

[0159] An embodiment of the present application also provides a computer program, which, when executed by a processor, implements the steps of the method for integrating and interacting modular knowledge graphs and retrieval-enhanced large models for financial branches in each embodiment of the present application.

[0160] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0161] In the above embodiments of the present application, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0162] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only exemplary. For example, the division of the units can be a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0163] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0164] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0165] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the existing technology, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk, etc. Various media that can store program code.

[0166] The above is only a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches, characterized by: include: Obtaining a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirement of the target object and a position vector for reflecting the position information of the target object in a financial branch; Determining, based on the position vector, a policy module accessible to the target object, wherein the policy module is a unit that logically divides policy documents according to the business areas of the financial branch; Determining a subgraph region corresponding to the question vector in a knowledge graph, wherein the knowledge graph is used to represent the logical structure between the policy documents; The document paragraph corresponding to the sub-graph area is retrieved from the sub-index library corresponding to the system module, and the document paragraph is input into the large model to generate a target answer corresponding to the query vector.

2. The method according to claim 1, characterized in that Determining the system module accessible to the target object based on the position vector includes: Obtain multi-dimensional tags corresponding to all policy modules, wherein the multi-dimensional tags include business domain tags for limiting the scope of application of the policy document, line attribute tags for refining the business lines under the business domain, and position applicability tags for limiting access rights; Determining the similarity between the position vector and the multidimensional label, and comparing the position authority information in the position vector with the position applicability label to obtain a comparison result; The policy modules accessible to the target object are determined based on the similarity and the comparison result, wherein each policy module corresponds to a sub-index library, and the sub-index library is used to store policy documents corresponding to the policy module.

3. The method according to claim 1, characterized in that The knowledge graph is constructed in the following way: Extracting entities from a set of policy documents, wherein the entities include policy clauses; Determine the relationship between the entities, and determine the edge of the knowledge graph based on the relationship, wherein the relationship includes one of the following: a reference relationship, an application relationship, an inheritance relationship, and a substitution relationship between the system clauses; The knowledge graph is constructed based on the entities and the edges.

4. The method according to claim 3, characterized in that Determining a subgraph region corresponding to the question vector in the knowledge graph includes: Matching the question vector with the knowledge graph to obtain a first graph node, wherein the first graph node includes a node whose similarity with the semantic feature of the question vector is greater than or equal to a first preset threshold; Searching for graph paths starting from the first graph node according to the edges of the knowledge graph, stopping the search when a stopping condition is met to obtain a second set of graph nodes, wherein the stopping condition includes that the similarity between the semantic features of the second graph node and the question vector is less than a second preset threshold, and the second preset threshold is less than the first preset threshold; The subgraph area is determined according to the first graph node and the second graph set.

5. The method according to claim 1, characterized in that Inputting the document paragraph into the large model to generate a target answer corresponding to the query vector includes: Obtaining guidance information for the large model, wherein the guidance information includes path information of the document paragraph in the knowledge graph and position information of the target object. The path information is used to indicate the institutional logic and structural order followed by the large model when generating answers; Constructing a target template according to the guidance information, wherein the target template includes control instructions corresponding to the path information and restriction conditions corresponding to the position information; The control instructions and the constraint conditions are processed using the large model to obtain the target answer.

6. The method according to claim 1, characterized in that After searching the sub-index library corresponding to the system module for the document paragraph corresponding to the sub-graph area, the method further includes: Obtaining metadata tags corresponding to the sub-nodes of the sub-graph area, wherein the metadata tags include the effective date, the abolition date, and the applicable region of the policy document; In the case where there are multiple versions of the policy document corresponding to the child node, determining a matching score between the policy document corresponding to each version and the query vector based on the metadata tag, wherein the matching score is used to quantitatively represent the applicability of the document paragraph corresponding to each version; A target document paragraph for generating the target answer is determined according to the matching score.

7. The method according to claim 6, characterized in that Determining a matching score between the policy document corresponding to each version and the query vector based on the metadata tag includes: Obtaining a query time corresponding to the query vector, and determining a first indicator based on the query time, the effective time, and the expiration time corresponding to the version, wherein the first indicator is used to quantitatively represent the timeliness of the version; Obtaining a query area corresponding to the query vector, and determining a second indicator based on the query area and the applicable area of the version, wherein the second indicator is used to quantitatively represent the regional compliance of the version; determining a third indicator between the question vector in the query vector and the version of the policy document, wherein the third indicator comprises semantic similarity; The matching score is determined based on the first indicator, the second indicator, and the third indicator.

8. A modular knowledge graph and retrieval-enhanced large model fusion interactive system for financial branches, characterized by: Including interactive equipment and servers, among which, The interactive device is connected to the server and is used to obtain a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirement of the target object and a position vector for reflecting the position information of the target object in a financial branch; The server is connected to the interactive device and is used to determine the policy module accessible to the target object based on the position vector, wherein the policy module is a unit that logically divides policy documents according to the business areas of the financial branch; determine the subgraph area corresponding to the question vector in the knowledge graph, wherein the knowledge graph is used to represent the logical structure between the policy documents; retrieve the document paragraph corresponding to the subgraph area in the sub-index library corresponding to the policy module, and input the document paragraph into the large model to generate a target answer corresponding to the query vector.

9. A modular knowledge graph and retrieval-enhanced large model fusion interaction device for financial branches, characterized by: include: an acquisition module, configured to acquire a query vector of a target object, wherein the query vector includes a question vector for expressing the query requirement of the target object and a position vector for reflecting the position information of the target object in a financial branch; a determination module, configured to determine, based on the position vector, a policy module accessible to the target object, wherein the policy module is a unit that logically divides policy documents according to the business areas of the financial branch; a matching module, configured to determine a subgraph region corresponding to the question vector in a knowledge graph, wherein the knowledge graph is configured to represent a logical structure between the policy documents; The interactive module is used to retrieve the document paragraph corresponding to the sub-graph area in the sub-index library corresponding to the system module, and input the document paragraph into the large model to generate a target answer corresponding to the query vector.

10. An electronic device, characterized in that: include: A memory and a processor, the memory being used to store program instructions; the processor being connected to the memory and being used to execute the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches as described in any one of claims 1 to 7.

11. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches as described in any one of claims 1 to 7 by running the computer program.

12. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by the processor, the modular knowledge graph and retrieval-enhanced large model fusion interaction method for financial branches as described in any one of claims 1 to 7 is implemented.

Images