Digital clone knowledge base construction and training method supporting multi-format document parsing

Abstract

The embodiment of the application provides a digital clone knowledge base construction and training method supporting multi-format document analysis, determines a target analysis path from a plurality of candidate analysis paths according to document features of a document to be processed, analyzes and generates a plurality of basic content units and structure information, and generates a source position address mapping capable of referring to a source document position for the basic content unit; performs segmentation, merging or aggregation on the basic content unit to construct an addressable semantic evidence unit, and performs knowledge extraction, cross-document alignment, version organization and index construction based on the same, to form an evidence binding knowledge base; and separates and trains a knowledge layer and a style layer of the digital clone based on the evidence binding knowledge base, to perform query or calling tasks, and systematically solves problems such as structure information loss, untraceable evidence, time state confusion, knowledge and style coupling, and high incremental update cost.

Description

Technical Field

[0001] This invention relates to the fields of intelligent document processing, knowledge organization, information retrieval, natural language processing, and digital avatar dialogue invocation, and particularly to a method for constructing and training a digital avatar knowledge base that supports multi-format document parsing, evidence addressing, time version management, controlled loading, and knowledge response package output, as well as corresponding systems, electronic devices, computer-readable storage media, and computer program products. Background Technology

[0002] With the deepening of personal knowledge management and organizational knowledge accumulation practices, and the rapid development of intelligent dialogue application technology, more and more users hope to build a digital avatar knowledge system that is highly consistent with their own experience and expression style by uploading materials such as resume documents, project documents, work logs, weekly reports, meeting minutes, debriefing reports, and achievement certificates, so as to drive the front-end digital avatar to present a consistent image and content in real-time dialogue.

[0003] Existing knowledge application solutions typically convert multi-format documents into plain text and segment them hierarchically according to fixed length or headings. They then construct semantic vectors and vector indexes to achieve text retrieval and answer generation. Alternatively, they may further build knowledge graphs, integrating graph retrieval with text retrieval to enhance the digital avatar's ability to answer complex questions. In addition, some solutions introduce features such as knowledge tracing, access control, and updates to improve project usability.

[0004] However, the inventors discovered that directly applying the above solution to the construction and management of the digital clone knowledge base still has the following obvious shortcomings: Loss of structural information: Directly converting documents to plain text can easily lead to the loss of page numbers, paragraph levels, table cell positions, list order, annotation paths, and cross-page structural information. In scenarios such as resumes, project experience, achievement tables, and meeting minutes, the same fact may appear across paragraphs, tables, or even pages. Dividing documents into hierarchical segments based on fixed lengths or headings can easily cause semantic breaks in events. When reviewing project experience or demonstrating job competency, this can easily lead to mismatches between responsibilities, achievements, and timelines.

[0005] Untraceable evidence: Converting documents to plain text for vector construction and indexing may allow for the retrieval of relevant content during subsequent calls, but it is difficult to accurately pinpoint the location of the original evidence. This can lead to substantial risks in scenarios such as resume generation, performance attribution, and audit Q&A, where the source of facts cannot be verified.

[0006] The problem lies in the inconsistent timeline: the representation of the same knowledge object in different documents, at different times, and in different versions often involves evolution, conflict, or supplementation. The above solutions lack a unified mechanism to handle time and version status and to arbitrate conflicts, making it difficult to reliably answer questions related to time or version status. Furthermore, in external presentations or internal audit scenarios, inconsistencies in timeline interpretation can easily lead to compliance risks.

[0007] Knowledge and style coupling: Many solutions model users’ factual knowledge and expression style in a mixed way, which leads to mutual contamination between the factual knowledge layer and style layer of the digital avatar. This results in situations where the answer style is similar but the facts are unreliable, or the facts are accurate but the answer style deviates, reducing the credibility of the digital avatar’s expression.

[0008] Incremental update costs are high: Existing solutions mostly involve one-time offline database creation. After receiving new documents, the entire database often needs to be re-splittered, embedded, and indexed. This results in high incremental costs and cannot meet the continuous update needs of the digital clone knowledge base. In scenarios where documents are frequently revised, this can easily lead to knowledge obsolescence and update delays.

[0009] Insufficient granularity of access control: Existing solutions typically output natural language text responses, which cannot output structured knowledge response objects that can be directly invoked by the digital avatar dialogue client and carry evidence addresses, access masks, version snapshots, and style tags. This makes it difficult for the front-end digital avatar to simultaneously adhere to knowledge evidence boundaries and access control boundaries during real-time dialogue.

[0010] Uncontrolled loading logic: Existing digital avatars usually focus on image consistency, action-driven or broadcast effects, and are separated from the evidence constraint interface of the private knowledge base. This can easily lead to the problem that the image is close, but the output content lacks personal knowledge basis.

[0011] Inconsistent front-end and back-end versions: The existing real-time dialogue links of digital avatars usually drive the front-end to broadcast the entire answer directly, lacking response grouping and session consistency identification for streaming dialogue; in low-latency scenarios, problems such as broadcasting before verification and inconsistency between the front-end version and the knowledge version in the same round of dialogue are likely to occur.

[0012] In summary, a new technical solution is urgently needed to address a series of problems in the construction and management of digital clone knowledge bases for multi-format documents, such as loss of structural information, untraceable evidence, chaotic time status, coupling of knowledge and style, insufficient granularity of access control, uncontrolled loading logic, inconsistency between front-end and back-end versions, and high cost of incremental updates. Summary of the Invention

[0013] This invention provides a method, system, and related products for constructing and training a digital clone knowledge base that supports parsing multiple document formats, in order to solve at least one of the following problems in the prior art: loss of structural information, untraceable evidence, chaotic time status, coupling of knowledge and style, insufficient granularity of permission control, uncontrolled loading logic, inconsistency between front-end and back-end versions, or high cost of incremental updates.

[0014] In a first aspect, embodiments of the present invention provide a method for constructing and training a digital clone knowledge base that supports multi-format document parsing, including: Receive at least one document to be processed, and perform format recognition, version identification, and preprocessing on the document to be processed; The preprocessed document to be processed is subjected to document feature extraction, and based on the document features, at least one target parsing path is determined from multiple candidate parsing paths; the multiple candidate parsing paths include at least two of the following: structural parsing path, layout parsing path, image parsing path, and hybrid parsing path; The document to be processed is parsed based on the target parsing path, generating multiple basic content units of the document to be processed, as well as the layout structure information or logical structure information corresponding to the basic content units. A source location address mapping that can point back to the source document location is generated for the basic content units. The source location address mapping is used for one or more of the following: evidence reference, evidence continuity judgment, version comparison, or answering source writing. Based on one or more of the following boundaries: content boundary, structural boundary, semantic boundary, temporal boundary, or event boundary, multiple basic content units are segmented, merged, or aggregated, and corresponding source location address mappings are attached to construct address-based semantic evidence units. Knowledge extraction, cross-document alignment, version organization, and index building are performed on semantic evidence units to form an evidence binding knowledge base. The evidence binding knowledge base includes one or more of the following: a set of semantic evidence units, knowledge objects associated with semantic evidence units, version organization structure, and retrieval index. Based on the evidence-binding knowledge base, the knowledge layer and style layer of the digital clone are trained separately, and the trained digital clone is used to perform tasks.

[0015] Secondly, embodiments of the present invention provide a digital clone knowledge base construction and training system that supports multi-format document parsing, including: The document access and version registration module is used to receive at least one document to be processed and to perform format recognition and preprocessing on the document to be processed. The parsing and routing module is used to extract document features from the preprocessed document to be processed, and determine at least one target parsing path from multiple candidate parsing paths based on the document features of the document to be processed; the multiple candidate parsing paths include one or more of the following: structure parsing path, layout parsing path, image parsing path, table parsing path, presentation parsing path, webpage parsing path, annotation parsing path, rule parsing path, model parsing path, or hybrid parsing path; The structure parsing and address mapping module is used to parse the document to be processed based on the target parsing path, generate multiple basic content units of the document to be processed, as well as the layout structure information or logical structure information corresponding to the basic content units, and generate source location address mappings for the basic content units that can point back to the source document location. The source location address mapping is used for one or more of the following: evidence reference, evidence continuity judgment, version comparison, or answer source writing back. The semantic evidence unit construction module is used to perform segmentation, merging or aggregation on multiple basic content units based on one or more of the following: content boundary, structural boundary, semantic boundary, time boundary or event boundary, and attach the corresponding source location address mapping to construct address-based semantic evidence units. The knowledge extraction and alignment module is used to perform knowledge extraction and cross-document alignment on semantic evidence units to obtain knowledge objects associated with semantic evidence units. The knowledge base organization and version indexing module is used to perform version organization and index construction on semantic evidence units to form an evidence binding knowledge base. The evidence binding knowledge base includes one or more of the following: a set of semantic evidence units, knowledge objects associated with semantic evidence units, version organization structure, and retrieval index. The hierarchical training module is used to separately train the knowledge layer and style layer of the digital clone based on the evidence-bound knowledge base, so as to run the trained digital clone to perform tasks. A storage cluster is used to store one or more of the following: basic content units, source location address mappings, semantic evidence units, knowledge objects, version organization structures, and retrieval indexes.

[0016] Thirdly, embodiments of the present invention provide an electronic device, including a memory, a processor, a communication interface, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the electronic device implements the above-described method or at least some of the functions of the above-described system.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, causes an electronic device to implement the above-described method or at least some of the functions of the above-described system.

[0018] Fifthly, embodiments of the present invention provide a computer program product, which includes a computer program. When the computer program is executed by a processor, it enables an electronic device to implement the above-described method or at least some of the functions of the above-described system.

[0019] The solution provided by the above-mentioned digital clone knowledge base construction and training method, system and related products that support multi-format document parsing systematically solves the problems of structural information loss, untraceable evidence, chaotic time state, knowledge and style coupling, insufficient granularity of permission control, uncontrolled loading logic, inconsistent front-end and back-end versions or high cost of incremental updates in related technologies by performing multi-path parsing, address mapping, boundary splitting / merging, knowledge extraction and alignment, version organization, and separation training of knowledge layer and style layer on the documents to be processed.

[0020] In this embodiment, multi-path parsing is used to extract the basic content units of the document, which can reduce the loss of structural information. A source location address mapping is generated for each basic content unit, so that the answer can accurately point back to the original evidence location. The basic content units are segmented, merged or aggregated based on content, structure, semantics, time or event boundaries to generate address-based semantic evidence units as the evidence basis. This can preserve the factual integrity across paragraphs and tables and avoid the mismatch of responsibilities, results and time caused by fixed segmentation. By performing knowledge object extraction and cross-document alignment on semantic evidence units and forming a version organization structure, different version states of the same knowledge object can be managed in a unified manner and support time version arbitration. By constructing an evidence binding knowledge base with semantic evidence units and knowledge objects as the core, the content output by the digital clone has personal knowledge basis and ensures the credibility of the content when working with the front-end image. Furthermore, by separating and training the knowledge layer and style layer of the digital clone, the fact selection and expression style are decoupled to avoid mutual contamination. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the structure of a digital clone knowledge base construction and training system supporting multi-format document parsing in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for constructing and training a digital clone knowledge base that supports multi-format document parsing, according to an embodiment of the present invention. Figure 3This is a schematic diagram illustrating the generation process of an address-based semantic evidence unit in one embodiment of the present invention; Figure 4 This is a schematic diagram of the response process to a query or call request in one embodiment of the present invention; Figure 5 This is a schematic diagram of the knowledge response packet generation and query request response process in one embodiment of the present invention; Figure 6 This is a schematic diagram of the incremental update process of the evidence binding knowledge base in one embodiment of the present invention; Figure 7 This is a schematic diagram of the logical data structure of a storage cluster in one embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. It should also be understood that, as used in this specification and the appended claims, the term "and / or" refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0025] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0028] To illustrate the technical solution of the present invention, specific embodiments are described below.

[0029] For ease of understanding, the following explanations are provided for the definitions of terms used in the embodiments of the present invention: A digital avatar is a knowledge-based digital entity that can be invoked by a dialogue system, formed based on the target user's historical documents, project records, resume materials, work experience, expression habits, and knowledge evidence. The digital avatar can work collaboratively with a front-end image presentation module that constructs image parameters from multiple reference images to achieve processes such as knowledge-based question answering.

[0030] A basic content unit refers to an atomic or quasi-atomic content object obtained after parsing the document to be processed by the target parsing path, which can be processed uniformly. It includes at least paragraph blocks, heading blocks, list items, tables, cells, fields, annotation blocks, text blocks corresponding to image areas, or combinations thereof. The parsing results output by different parsing paths are normalized and mapped to basic content units with a unified structure to facilitate subsequent reading order reconstruction, source location address mapping, and address-based semantic evidence unit construction.

[0031] Source location address mapping refers to a set of address information that points the processed content unit, evidence unit, or response fragment back to the corresponding location in the source document. Source location address mapping can be implemented using physical page addresses, logical structure addresses, character offset addresses, table cell addresses, comment addresses, attachment paths, image region addresses, or combinations thereof. For text, web pages, structured tables, or other formatted documents without fixed page coordinates, this source location address mapping can be implemented using logical structure paths, character offsets, field paths, or content block numbers.

[0032] Address-based semantic evidence units refer to knowledge-bearing units obtained from parsing source documents that possess both semantic integrity and the ability to refer back to the source location. The source location can be any one of the following: physical layout location, logical structure location, character offset location, table or cell location, annotation location, attachment location, or image area location, or a combination of multiple locations. Not all documents are required to have page coordinates.

[0033] Evidence-binding fact graph refers to a graph structure with people, organizations, projects, tasks, results, skills, events, time, etc. as nodes, and relationships such as participation, responsibility, output, occurrence, proof, and dependence as edges, where each node and relation edge is bound to at least one addressable semantic evidence unit and its address set.

[0034] A time version chain refers to a chain-like or directed acyclic structure that organizes the version status of the same knowledge object under different document versions, different generation times, and different effective times. It is used to support time replay and version arbitration.

[0035] Knowledge layer training refers to training on fact selection, evidence matching, version consistency, citation location, and evidence coverage.

[0036] Style layer training refers to training that targets users' terminology preferences, sentence structure preferences, expression length, narrative perspective, and organizational habits.

[0037] A knowledge response package refers to a structured response object that the system outputs to the digital avatar dialogue client after a query. It is not limited to the text answer itself, but includes at least several items such as the answer text, evidence address set, permission mask, version snapshot, style tags, and response confidence.

[0038] Conversation consistency identifiers refer to the identifiers that uniformly constrain the front-end image version, knowledge version, and style version within a single round of dialogue, in order to avoid version inconsistencies in the same round of output.

[0039] A response group refers to a sub-response object that is sequentially streamed and divided from a knowledge response package. It includes at least several of the following: response fragment, evidence address subset, permission mask subset, version snapshot, style tag, and time sequence number.

[0040] To address the problems of structural information loss, untraceable evidence, chaotic temporal status, knowledge and style coupling, insufficient granularity of permission control, insufficient controlled loading capability, and high incremental update cost in the construction, training, and dialogue invocation of multi-format documents in digital avatar knowledge bases, this invention provides a highly feasible, traceable evidence-based, knowledge and style-separable training, and efficient incremental update method for constructing and training digital avatar knowledge bases. Furthermore, the knowledge response information generated based on the constructed or updated evidence-bound knowledge base can be directly invoked by the digital avatar dialogue client, improving the efficiency of knowledge invocation and response output.

[0041] Specifically, after receiving the document to be processed, the evidence binding knowledge base is constructed and updated in the following way: The document to be processed undergoes format recognition and preprocessing in order to register the document and obtain document registration records; When the document to be processed is a new topic document, document features are extracted from the document to be processed, and at least one target parsing path is determined from multiple candidate parsing paths based on the document features of the document to be processed. The document to be processed is parsed based on the target parsing path to generate multiple basic content units of the document to be processed, as well as the layout structure information or logical structure information corresponding to the basic content units. Based on the basic content units and the corresponding page layout or logical structure information, the reading order of the document to be processed is reconstructed to obtain the reading order number of at least some of the basic content units. Based on document registration records and reading sequence numbers, a source location address mapping is generated for the basic content unit that can point back to the source document location. The source location address mapping includes the physical address, logical address, or a combination of both of the basic content unit. The source location address mapping is used for one or more of the following: evidence reference, evidence continuity judgment, version comparison, or answering source writing back. Based on one or more of the following boundaries: content boundary, structural boundary, semantic boundary, temporal boundary, or event boundary, multiple basic content units are segmented, merged, or aggregated, and corresponding source location address mappings are attached to construct address-based semantic evidence units. Perform one or more of the following on semantic evidence units: knowledge extraction, cross-document alignment, version organization, or index building, to build an evidence binding knowledge base. The evidence binding knowledge base includes one or more of the following: a set of semantic evidence units, knowledge objects associated with the semantic evidence units, a version organization structure, and a retrieval index. After the evidence binding knowledge base is formed, when a new document, updated document, deletion instruction or permission change information is received, the change set is identified based on the hierarchical fingerprint of the involved object, and differential reconstruction, update or invalidation processing is performed on one or more of the semantic evidence unit, knowledge object, version status, permission tag or index item corresponding to the change set to form a new version of the evidence binding knowledge base to train the digital clone.

[0042] Among them, after building or updating the evidence binding knowledge base, one or more of the following are performed on the digital clone: ​​training, parameter calibration, configuration or invocation strategy optimization, based on the evidence binding knowledge base. During the training process, the knowledge layer and style layer of the digital clone are trained separately so as to run the trained digital clone to perform the task. During the operation of the digital clone, when a query or call request is received, query constraint information is generated, which includes one or more of the following: time constraints, permission constraints, evidence constraints, style constraints, context budget constraints, and latency budget constraints. Based on the query constraint information, candidate filtering, recall sorting, controlled loading, and permission pruning are performed on multiple semantic evidence units to obtain a loaded evidence set. A knowledge response package is generated based on the loaded evidence set, and according to the evidence coverage status, permission status, version status, or confidence status of the knowledge response package, one of the following is output: normal response, downgraded response, rejection prompt, or pending review prompt.

[0043] Compared with existing technologies, this invention does not simply convert multi-format documents into plain text for vector retrieval. Instead, it enables source location address mapping to participate simultaneously in semantic evidence unit construction, controlled loading, and response source write-back; enables temporal version chains to participate simultaneously in candidate pre-filtering, version arbitration, and response confidence calculation; enables the knowledge response package to carry response text, evidence address, permission mask, and version snapshot simultaneously, and performs hierarchical output based on response confidence and evidence coverage; and further supports real-time dialogue invocation of a digital avatar front-end built from multiple reference images through session consistency identification and response grouping mechanisms.

[0044] The technical solutions provided by the embodiments of the present invention have at least the following beneficial effects: Firstly, by using address-based semantic evidence units, the response can be traced back to the page number, paragraph, table, or cell location of the original document, thereby improving the traceability of evidence and the verifiability of the response. Secondly, semantic evidence units are constructed by jointly using page boundaries, semantic boundaries, and event boundaries, which reduces semantic breaks caused by fixed-length segmentation and improves the stability of fact extraction, event extraction, and retrieval ranking. Third, by binding evidence to fact graphs and version organization structures (such as time version chains), it is possible to arbitrate the status of the same knowledge object at different times and in different versions, thereby improving the accuracy of time-sensitive question answering; Fourth, by decoupling the knowledge layer training from the style layer training, the stability of fact selection is ensured while user expression habits are preserved, and the interference of style transfer on the fact layer is reduced. Fifth, by using query constraint parsing, candidate pre-filtering, hybrid recall, controlled loading, and knowledge response package generation, evidence relevance, permission boundaries, context budget, and response latency can be uniformly controlled, thereby improving the controllability of the call; Sixth, by controlling the output status through response confidence and evidence coverage, hierarchical degradation is performed when there is insufficient evidence, insufficient permissions, or version conflicts, thereby reducing the risk of output without evidence and output without authority. Seventh, differential reconstruction is performed based on hierarchical fingerprints, and only the semantic evidence units, node edges and index items of the fact graph corresponding to the change set are reconstructed, thereby reducing the cost of continuous updates; Eighth, in some embodiments, address back pointer accuracy, evidence coverage, permission pruning error rate, response latency exceeding limit rate, and differential reconstruction ratio can be used as indicators for quantitative evaluation or parameter calibration, and corresponding parameter values ​​can be configured according to application scenarios, document types, and permission levels, thereby facilitating continuous evaluation and online optimization of system quality.

[0045] The method for constructing and training a digital clone knowledge base that supports multi-format document parsing provided in this invention can be applied to, for example... Figure 1 The digital clone knowledge base construction and training system (hereinafter referred to as the system) that supports multi-format document parsing communicates with user terminals through the network to realize document reception and input of related instructions.

[0046] This system can be a server or a terminal device with computing capabilities. The user terminal refers to the terminal device held by the user, including but not limited to various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0047] In practical applications, users send one or more documents to be processed to the system through user terminals. After receiving the documents to be processed from the user terminals, the system performs format recognition and corresponding knowledge processing on the documents to be processed to generate multiple address-based semantic evidence units and build or update the evidence binding knowledge base. Then, based on the built or updated evidence binding knowledge base, the system trains, calibrates, configures, and optimizes the calling strategy of the digital clone so that the digital clone can be called to perform the corresponding tasks.

[0048] Furthermore, the key to this invention's embodiments lies not in simply listing document parsing, knowledge extraction, version management, response generation, and digital avatar invocation in parallel, but in: ensuring that source location address mapping permeates basic content units, semantic evidence units, controlled loading, and response source write-back; ensuring that version organization structure permeates candidate pre-filtering, version arbitration, and response control; ensuring that hierarchical fingerprints permeate change set identification, linked reconstruction, and index updates; and ensuring that session consistency identifiers permeate response group output and front-end dialogue invocation. Thus, while improving evidence traceability and version consistency, it also reduces the processing overhead caused by full database re-splitting, full re-embedding, and full re-indexing, thereby improving the stability and real-time performance of query responses.

[0049] Specifically, such as Figure 1 As shown, the system includes a document access and version registration module, a parsing and routing module, a structure parsing and address mapping module, a semantic evidence unit construction module, a knowledge extraction and alignment module, a knowledge base organization and version indexing module, a hierarchical training module, a query orchestration module, a candidate pre-filtering and controlled loading module, a knowledge response package generation module, a permission and security control module, an incremental update and evaluation module, a digital avatar dialogue interface module, and a storage cluster. The digital avatar dialogue interface module is used to communicate with the digital avatar dialogue caller, the knowledge question answering caller, and the knowledge service interface to output knowledge response packages or response groups. The functional modules in the system are defined as follows: The document access and version registration module is used to receive at least one document to be processed, perform format recognition and preprocessing on the document to be processed, and register the document to obtain a document registration record; The parsing and routing module is used to determine at least one target parsing path from multiple candidate parsing paths based on the document characteristics of the document to be processed when the document to be processed is a new topic document. The structure parsing and address mapping module is used to parse the document to be processed based on the target parsing path, generate multiple basic content units of the document to be processed, as well as the layout structure information or logical structure information corresponding to the basic content units, and perform reading order reconstruction on the document to be processed based on the basic content units and the corresponding layout structure information or logical structure information to obtain the reading order sequence number of at least some of the basic content units. The structure parsing and address mapping module is also used to generate source location address mappings that can point back to the source document location for basic content units based on document registration records and reading order numbers. The source location address mapping includes the physical address, logical address, or a combination of both of the basic content units. The source location address mapping is used for one or more of the following: evidence reference, evidence continuity judgment, version comparison, or answering source write-back. The semantic evidence unit construction module is used to perform segmentation, merging or aggregation on multiple basic content units based on one or more of the following: content boundary, structural boundary, semantic boundary, time boundary or event boundary, and attach the corresponding source location address mapping to construct address-based semantic evidence units. The knowledge extraction and alignment module is used to perform knowledge extraction and cross-document alignment on semantic evidence units to obtain knowledge objects associated with semantic evidence units. The knowledge base organization and version indexing module is used to perform version organization and index construction on semantic evidence units to form an evidence binding knowledge base. The evidence binding knowledge base includes one or more of the following: a set of semantic evidence units, knowledge objects associated with semantic evidence units, version organization structure, and retrieval index. The layered training module is used to construct training samples for semantic evidence units in the evidence binding knowledge base, and to perform one or more of the following on the digital clone: ​​training, parameter calibration, configuration or invocation strategy optimization. During the training process, the knowledge layer and style layer of the digital clone are trained separately so that the trained digital clone can perform the task. The query orchestration module is used to generate query constraint information based on the query or call request when a query or call request is received during the operation of the digital clone. The query constraint information includes one or more of the following: time constraint, permission constraint, evidence constraint, style constraint, context budget constraint, and latency budget constraint. The candidate pre-filtering and controlled loading module is used to perform candidate filtering, recall sorting and controlled loading on multiple semantic evidence units based on query constraint information to obtain a loaded evidence set. The knowledge response package generation module is used to generate a knowledge response package based on the loaded evidence set; The permission and security control module is used to perform permission pruning on semantic evidence units based on query constraint information when generating and loading evidence sets or knowledge response packages; The digital avatar dialogue interface module is used to output one of the following responses based on the evidence coverage status, permission status, version status, or confidence status of the knowledge response package: normal response, downgraded response, rejection prompt, or pending review prompt. The incremental update and evaluation module is used to, after the evidence binding knowledge base is formed, when receiving new documents, updated documents, deletion instructions or permission change information, identify the set of changes based on the hierarchical fingerprint of the objects involved, and perform differential reconstruction, update or invalidation processing on one or more of the semantic evidence units, knowledge objects, version status, permission tags or index items corresponding to the set of changes, to form a new version of the evidence binding knowledge base to train the digital clone; the changed documents include new documents and updated documents.

[0050] The storage cluster supports read and write operations for the document access and version registration module, storage structure parsing and address mapping module, semantic evidence unit construction module, incremental update and evaluation module, knowledge extraction and alignment module, and knowledge base organization and version indexing module. It stores information generated by the modules, such as the generated basic content units and their source location address mappings, as well as semantic evidence units, knowledge base organization, and evidence binding knowledge bases.

[0051] This system, through the coordinated operation of its various functional modules, enables the construction, training, updating, and response of a digital avatar knowledge base supporting the parsing of multi-format documents. By performing format recognition, parsing path selection, layout structure generation, and source location address mapping on multi-format documents, it constructs semantic evidence units that can refer back to the source documents, forming an evidence-bound knowledge base, a time-version chain, and a hybrid index. During the training phase, the knowledge layer and style layer are separated. During the query phase, pre-filtering, recall sorting, controlled loading, and knowledge response package generation are completed based on time, permissions, evidence, style, and budget constraints. When documents are updated, hierarchical fingerprint differential reconstruction improves evidence traceability, version consistency, controllable calls, and continuous update capabilities. This systematically solves problems in existing technologies such as lost structural information, untraceable evidence, chaotic time states, coupling of knowledge and style, insufficient granularity of permission control, uncontrolled loading logic, inconsistent front-end and back-end versions, or high costs of incremental updates.

[0052] Furthermore, the evidence-binding knowledge base formed in this embodiment includes a set of semantic evidence units, knowledge objects, a version organization structure, and a retrieval index. It can perform differential updates based on hierarchical fingerprints, avoiding full reconstruction. Through the version organization structure and address mapping, evidence of the corresponding version can be quickly selected to generate a directly output knowledge response package, which can be directly invoked by the digital avatar dialogue client. This package carries structured knowledge response objects such as evidence addresses, permission masks, version snapshots, and style tags, improving the granularity of permission control and ensuring that the front-end digital avatar can synchronously adhere to knowledge evidence boundaries and permission control boundaries during real-time dialogue. The version organization structure and address mapping provide a foundation for subsequent response grouping and session consistency identification, ensuring that the front-end version and knowledge version remain consistent within the same round of dialogue.

[0053] In one embodiment, such as Figure 2 As shown, a method for constructing and training a digital clone knowledge base that supports multi-format document parsing is provided, and this method is applied to... Figure 1 Taking the system in the example, the following steps S11 to S28 are described. The specific implementation process of each step is as follows: S11: Receive at least one document to be processed.

[0054] Users send at least one document to be processed to the system via their user terminals. This document can be an initial document used to build the knowledge base, or a modified document used to update the knowledge base. Modified documents include newly added documents (i.e., new topic documents) and updated documents (i.e., new versions of existing documents). The system's document access and version registration module receives the documents to be processed sent by the user terminal.

[0055] During the knowledge base construction phase, the system executes the knowledge base construction path based on the received documents to be processed, i.e., steps S12 to S18, to construct the evidence binding knowledge base. During the knowledge base update phase, the system executes an incremental update path based on the received documents to be processed. The incremental update path refers to steps S12 to S18 to reconstruct the changed addressable semantic evidence units, and then performs knowledge extraction, cross-document alignment, version organization, and index construction to update the evidence binding knowledge base to a new version.

[0056] After building and updating the evidence binding knowledge base, the digital clone is trained and invoked based on the latest version of the evidence binding knowledge base. During the process of creating the digital clone, parameter calibration and online updates are performed, as well as configuration and strategy optimization operations are performed.

[0057] S12: Perform format recognition, version identification, and preprocessing on the document to be processed to determine the type of the document and register it to obtain a document registration record.

[0058] After receiving the document to be processed sent by the user terminal, the document access and version registration module performs format recognition, identifier generation, fingerprint generation and preprocessing on the document to identify the type of the document to be processed, and performs version registration based on the generated identifier, document fingerprint and preprocessing results, thereby obtaining the document registration record.

[0059] The document registration record includes at least the document format, document fingerprint, document identifier, topic identifier, version identifier, parent version identifier, source channel, creation time, permission information (such as permission level), parsing path, and processing status. Document formats include editable formats and fixed layouts. The parent version identifier for new topic documents is empty. The parsing path is written to the document registration record after the target parsing path is determined. The processing status identifies the current stage of the document to be processed, including registration status, parsing status, basic content unit construction status, address mapping status, semantic evidence unit construction status, and completion status.

[0060] The document access and version registration module can generate a document fingerprint for the document to be processed. This fingerprint generation process includes: obtaining the metadata and structural summary of the document to be processed. The metadata includes one or more of the following: filename, timestamp, source channel, and author identifier. The structural summary includes one or more of the following: page number, paragraph number, table number, style number, and total character count. The original byte content, metadata, and structural summary of the document to be processed are concatenated, and a hash function is used to calculate the hash value of the concatenated result to obtain the document fingerprint. The document fingerprint can be calculated using the following formula: ; in, Represents a document fingerprint; Represents a hash function; Indicates a splicing operation; This represents the raw byte content of the document to be processed; Represents the metadata of the document to be processed; This represents a structural summary of the document to be processed.

[0061] The preprocessing includes one or more of the following: source file integrity verification, encoding normalization, format standardization, security scanning, topic identification, metadata standardization, parent version identification, initial permission marking, abnormal content detection, and de-identification detection, as well as duplicate detection.

[0062] Specifically, the document access and version registration module can first perform source file integrity verification on the document to be processed. Simultaneously, it can perform security scanning, abnormal content detection, and desensitization detection to filter malicious or corrupted documents. If the document fails the source file integrity verification, security scanning, abnormal content detection, or desensitization detection, it returns a corresponding file insufficiency warning, such as a file incomplete warning, a file insecure warning, a file content abnormality warning, or a file containing sensitive information warning, prompting the user to modify or resend the file. After the document passes the source file integrity verification, security scanning, abnormal content detection, and desensitization detection, it undergoes encoding normalization (i.e., unified character encoding) and format standardization (i.e., conversion to a preset standard format) to obtain a standardized document for subsequent processing, ensuring that the document can be parsed correctly later.

[0063] Specifically, topic recognition identifies the topic of the document to be processed and generates topic tags for subsequent candidate filtering. Metadata standardization standardizes the metadata of the document to be processed, unifying fields such as filename, author, timestamp, and source channel, facilitating subsequent version chain arbitration and permission inheritance. Initial permission tagging sets an initial permission level for at least one of the document, paragraph, or field to be processed, generating corresponding permission information for subsequent permission masks and penalties for permission violations during queries.

[0064] Specifically, duplicate detection can be used to detect duplicate content in a document to be processed, and / or to detect whether the document to be processed is a new topic document, and to perform different version management based on the detection results. For example, the document to be processed can be matched with existing historical documents by document topic matching, document fingerprint matching, or structural path matching to calculate at least one or all of the topic matching degree, fingerprint matching degree, and structural path matching degree. Based on at least one or all of the topic matching degree, fingerprint matching degree, and structural path matching degree, it can be determined whether the document to be processed is a new topic document, i.e., a newly added document. When the corresponding matching degree is less than a preset matching degree threshold, the document to be processed is determined to be a new topic document; otherwise, when the corresponding matching degree is greater than or equal to the preset matching degree threshold, the document to be processed is determined to be a new version of an existing document, i.e., an updated document.

[0065] Each matching degree can be calculated using similarity. The structural path is the hierarchical location path of content within the document, including structural information such as outlines, table of contents, heading levels, or page layout. The threshold values ​​for topic matching are 0.70–0.95, for structural path matching are 0.55–0.90, and for fingerprint matching are 0.75–0.95.

[0066] In one embodiment, when the topic matching degree is less than the topic matching threshold, or the fingerprint matching degree is less than the fingerprint matching degree threshold, or the structure path matching degree is less than the structure path matching degree threshold, the document to be processed is a new topic document. When the topic matching degree is greater than or equal to the topic matching degree threshold, and the fingerprint matching degree is greater than or equal to the fingerprint matching degree threshold, and the structure path matching degree is greater than or equal to the structure path matching degree threshold, the document to be processed is determined to be a new version of an existing document, rather than a new topic document. By using a multi-dimensional joint determination based on topic matching degree, fingerprint matching degree, and structure path matching degree, the accuracy of determining new topic documents is improved.

[0067] For new topic documents, they are treated as independent documents with newly created document identifiers, topic identifiers, and version identifiers. Version relationships from existing documents are not inherited. Subsequent parsing and evidence construction processes are performed as if the document were entirely new, resulting in newly added semantic evidence units and incremental updates. For new versions of existing documents, they can be registered as new versions of the same document. Parent version identification is performed to inherit the parent version's document identifier and topic identifier. The parent version identifier is then associated and included in the time version chain management. Subsequent parsing and evidence construction will be linked to the parent version to obtain newly added or modified semantic evidence units and perform incremental updates.

[0068] S13: Extract the document features of the document to be processed, and select at least one target parsing path from multiple candidate parsing paths based on the document features.

[0069] The parsing and routing module obtains the preprocessed document to be processed, calculates its document features, and determines one or more target parsing paths from multiple candidate parsing paths based on these features. The document features are used to select the target parsing path that can generate basic content units, but the basic content units are not limited to being generated by only one parsing path.

[0070] The document features are used to characterize the parsability, structure, or layout complexity of the document to be processed. Specifically, the document features may include one or more of the following: raw text extractability rate, image ratio, table density, multi-column probability, layout complexity, style hierarchy strength, coding reliability, formula density, number of in-page blocks, logical structure strength, footnote density, annotation density, number of attachments, document source information, and historical parsing results.

[0071] Specifically, the multiple candidate parsing paths include one or more of the following: structural parsing path, layout parsing path, image parsing path, table parsing path, presentation parsing path, webpage parsing path, annotation parsing path, hybrid parsing path, rule parsing path, and model parsing path. The structural parsing and address mapping module can determine at least one target parsing path from among the multiple candidate parsing paths that can output a unified basic content unit, based on the aforementioned document features. Here, the basic content unit refers to an atomic or quasi-atomic content object obtained by the target parsing path after parsing the document to be processed, which can be processed uniformly; this basic content unit includes at least one or a combination of at least two of the following: paragraph blocks, heading blocks, list blocks, table blocks, cell blocks, field blocks, annotation blocks, and text blocks corresponding to image areas in the document. By selecting a target parsing path that can output a unified basic content unit based on document features, adaptive parsing of the document to be processed is achieved, significantly improving the parsing adaptability to documents of different formats and complexities, and the standardization of basic content units.

[0072] Specifically, based on the document format and the aforementioned document characteristics, a target parsing path can be selected from multiple candidate parsing paths according to preset rules. For example, when the document to be processed is in an editable format and the style hierarchy strength is greater than the preset strength (i.e., the style hierarchy is obvious), the structure parsing path can be selected as the target parsing path; when the document to be processed has a fixed layout and the native text extraction rate is greater than or equal to the preset extraction rate, the layout parsing path can be selected as the target parsing path.

[0073] Furthermore, when the document to be processed is determined to have a high degree of scannability based on one or more of the following factors: native text extractability, image representation ratio, encoding reliability, and style hierarchy strength, the image parsing path can be selected as the target parsing path. If the native text extractability is less than a preset extraction rate, the image representation ratio is greater than a preset ratio, and the encoding reliability is less than a preset reliability, the scannability is determined to be high; conversely, the scannability is low. When the document to be processed is determined to contain both native text pages and image pages based on the native text extractability, image representation ratio, encoding reliability, and layout complexity, a hybrid parsing path can be selected as the target parsing path.

[0074] In one embodiment, the parsing routing module can score each candidate parsing path based on the aforementioned document features to obtain a score value for each candidate parsing path. Based on at least one of the score values, computational costs, parsing confidence, or historical parsing performance of each candidate parsing path, at least one target parsing path capable of outputting a unified basic content unit is determined from among the multiple candidate parsing paths. For example, the candidate parsing path with the highest score value can be selected as the target parsing path, or the candidate parsing path with the highest parsing confidence or the best historical parsing performance can be selected as the target parsing path.

[0075] The score can be calculated using a linear weighted model, a rule-based model, a classification model, or a combination thereof, and can be obtained by weighted summation of multiple document features based on the feature weights of the candidate parsing path. The parsing confidence score can be a path reliability score based on historical data statistics and labeling, such as the parsing success rate of the candidate parsing path for documents of the same topic and / or format. The historical parsing performance can be calculated based on the historical parsing performance (such as parsing quality and error rate) of the candidate parsing path in documents of the same topic and / or format.

[0076] In one embodiment, the routing resolution module can calculate the total score of a candidate resolution path based on at least two of the following: a score value, computational cost, resolution confidence, or historical resolution performance (represented by a normalized numerical value). Among multiple candidate resolution paths, the candidate resolution path with the highest total score is selected as the target resolution path. For example, the computational cost of a candidate resolution path can be multiplied by a penalty coefficient to obtain a penalty, and the total score of the candidate resolution path can be obtained by subtracting the penalty from the score value. This total score can be calculated using the following formula: ; in, Indicates the target parsing path; Indicates the first Document characteristics; Indicates candidate parsing paths The corresponding feature weights; This represents the score of the candidate parsing path; Indicates candidate parsing paths The computational cost is a value pre-calibrated based on historical or experimental data; This represents the penalty coefficient for calculating the cost; it is a pre-calibrated constant. This represents the total number of document features.

[0077] In other embodiments, the total score of the candidate parsing path can be calculated in other ways, such as by weighting and summing the score value, parsing confidence, or historical parsing effect of the candidate parsing path, and then subtracting the cost penalty to obtain the total score of the candidate parsing path.

[0078] S14: Based on the target parsing path, parse the document to be processed to generate multiple basic content units of the document to be processed, as well as the layout structure information or logical structure information corresponding to the basic content units.

[0079] After determining the target parsing path, the structure parsing and address mapping module parses the document to be processed based on the target parsing path to extract paragraphs, headings, lists, tables, cells, fields, annotations, image regions, and corresponding text. After normalization, these are output as page blocks in a uniform format as basic content units. Simultaneously, based on the structure and position of each basic content unit in the document, the corresponding page structure information and / or logical structure information are determined. The parsing results output by the target parsing path are normalized and mapped to basic content units with a uniform structure, facilitating subsequent reading order reconstruction, source location address mapping, and construction of address-based semantic evidence units.

[0080] One parsing path can correspond to one decoder. The structure parsing and address mapping module can call the decoder corresponding to the target parsing path to perform multi-class parsing of the document to be processed, and obtain multiple basic content units, as well as their corresponding layout structure information and / or logical structure information.

[0081] The layout structure information reflects the physical layout of basic content units in the document. This layout structure information includes the type of basic content unit (i.e., block type), the page it is located on (which can be indicated by the page number), the paragraph it is located on (which can be indicated by the paragraph number), the list it is in, the table it is in, the cell it is in, and its coordinates, etc. The logical structure information reflects the content hierarchy and semantic information of basic content units in the document. This logical structure information includes the page number of the basic content unit, the chapter title level, the list nesting sequence number, the paragraph logical affiliation, and the structural path (such as chapter / paragraph), etc.

[0082] In one embodiment, during the parsing of the document to be processed based on the target parsing document, the confidence level of various parsing results is calculated, such as the confidence level of text parsing, table parsing, layout parsing, image parsing, or mixed parsing. When the confidence level of any of the parsing results is lower than a preset confidence threshold, it indicates that the parsing result is not accurate enough. This can trigger one or more of the following processes: backtracking, supplementary parsing, or manual review. This effectively ensures the accuracy and reliability of the parsing results and provides high-quality basic data for the subsequent construction of semantic evidence units and the generation of the knowledge base.

[0083] S15: Based on the basic content units and the corresponding page layout or logical structure information, perform reading order reconstruction on the document to be processed to obtain the reading order number of at least some of the basic content units.

[0084] The structure parsing and address mapping module reconstructs the reading order of the document to be processed based on the basic content units and the corresponding page structure information and / or logical structure information. This allows the module to determine the logical order in which the user reads each basic content unit based on its position in the document, thereby obtaining the reading order number of at least some or all of the basic content units.

[0085] In one embodiment, the reading order reconstruction includes: determining the relative vertical position, column consistency, style continuity, sequence number continuity, and overlapping relationship between basic content units based on basic content units and page layout or logical structure information; and determining the reading order between basic content units based on one or more of the relative position, column consistency, style continuity, sequence number continuity, and overlapping relationship between basic content units, thereby determining the reading order number of the basic content units. This achieves accurate restoration of the document's true reading order and provides a reliable sequence basis for subsequent address mapping, semantic evidence unit construction, and continuous loading.

[0086] Specifically, based on the basic content units and their corresponding page layout and / or logical structure information, one or more of the following relationships are determined between the basic content units: relative position, column consistency, style continuity, sequence number continuity, and overlap. This establishes directed relationships between the basic content units, which indicate the reading order between two basic content units. The basic content units are then hierarchically organized according to these directed relationships to generate a candidate page layout tree for the document to be processed. Specifically, basic content units such as paragraph blocks, heading blocks, list blocks, table blocks, cell blocks, field blocks, annotation blocks, or text blocks corresponding to image areas can be used as nodes. The directed relationships between two nodes that meet the specified conditions are used as edges between the two nodes, forming a directed graph, which yields the candidate page layout tree. The candidate page layout tree may contain multiple paths (i.e., reading order), and may even have loops. To obtain a unique linear reading order, a preset topological sorting algorithm, an approximate topological sorting algorithm, or a bundle search algorithm with a penalty term can be used to perform topological sorting or path search on each node in the candidate page layout tree to obtain a page layout tree with a unique path. Based on the node connection order in the page layout tree, each node is labeled with a reading order number to obtain the reading order number of each node, which is the reading order number of each basic content unit.

[0087] Specifically, a directed reading connection weight can be constructed between any two basic content units to represent the directed relationship between the two corresponding nodes. When the reading connection weight is greater than a preset weight, it indicates that there is a directed relationship between the two basic content units and the condition is met, thus constructing an edge between the two basic content units. Specifically, based on the layout structure information and / or logical structure information corresponding to the two basic content units, the relative vertical position, column consistency, style continuity, sequence number continuity, and overlap relationship between the two basic content units can be determined to calculate the relative vertical position score, column consistency score, style continuity score, sequence number continuity score, and overlap penalty for the two basic content units. A weighted sum is then performed based on the relative vertical position score, column consistency score, style continuity score, and sequence number continuity score, and the overlap penalty is subtracted from the weighted sum to obtain the reading connection weight between the two basic content units, i.e., the reading connection weight between the two layout blocks. This reading connection weight is a directed weight used to represent the reading order of the two layout blocks.

[0088] The relative vertical position indicates the vertical relationship between two basic content units. If page block A is above page block B, and A is read first, then the relative vertical position of page blocks A and B is scored. Column consistency indicates whether the two basic content units are in the same column. Style continuity indicates the similarity of styles such as font, font size, and indentation. If the styles are the same or similar, it indicates that they may be the same paragraph or continuous content. Sequence continuity indicates the sequence of the numbers of the two basic content units. Overlap indicates whether the coordinates of the two basic content units overlap. Points are deducted when there is overlap to avoid misjudging it as a sequential relationship.

[0089] The reading connection weight between two page blocks can be calculated using the following formula: ; in, Indicates from page block To the page Reading connection weight; The score is based on the relative position of the top and bottom. Indicates the column consistency score; Indicates the score for style continuity; Indicates the score for the continuity of serial numbers; Indicates overlapping penalties; This represents the pre-configured weighting coefficients.

[0090] S16: Based on the document registration record and reading sequence number, generate a source location address mapping for the basic content unit that can point back to the source document location. The source location address mapping includes the physical address, logical address, or a combination of both of the basic content unit.

[0091] The structure parsing and address mapping module generates a source location address mapping for each basic content unit based on the document identifier, version identifier, reading order number, corresponding page structure information, and / or logical structure information from the document registration record. This source location address mapping is used for one or more of the following: evidence referencing, evidence continuity judgment, version comparison, or answering source rewriting.

[0092] Source location address mapping can be a physical address based on page coordinates, or a logical address pointing to chapters, paragraphs, tables, fields, comments, image regions, or character offsets, or a combination of physical and logical addresses. As a result, the system can be compatible with different types of source documents, such as PDFs, word processing documents, web pages, plain text, exported text from tables, exported text from presentations, and image documents.

[0093] In one embodiment, the source location address mapping includes a document identifier of the document to which the basic content unit belongs (i.e., the document to be processed), and location information for locating the original content location (i.e., the location of the basic content unit in the document to be processed). This location information includes one or more of the following: the version identifier of the document to which the basic content unit belongs, the content unit identifier, the page number or logical page number, normalized coordinates, the reading order number, the content unit type (i.e., the block type), the character start and end offsets, the structure path, the field path, the table path, the cell path, the comment path, the attachment path, and the image region path.

[0094] The normalized coordinates are calculated by normalizing the original coordinates of the basic content unit based on the page size of the page containing the basic content unit. This ensures that the address mapping is independent of the absolute size of the original page, facilitating proportional positioning in different rendering environments. The character start and end offsets are determined based on the start and end character positions of the basic content unit within the entire document text flow, representing the positional offset of the start and end characters for precise character-level positioning. The structural path can be generated during parsing or based on the layout tree, representing the hierarchical path from the document root node to the basic content unit.

[0095] In other embodiments, after obtaining the basic content unit and its corresponding layout structure information and / or logical structure information, it is not necessary to perform reading order reconstruction. Instead, the position of the basic content unit in the document can be generated based on the layout structure information and / or logical structure information, and its source position address mapping can be constructed.

[0096] S17: Based on one or more of the following: content boundaries, structural boundaries, semantic boundaries, temporal boundaries, or event boundaries, perform segmentation, merging, or aggregation on multiple basic content units and attach corresponding source location address mappings to construct address-based semantic evidence units.

[0097] The semantic evidence unit construction module takes basic content units as input and integrates one or more of content boundaries, structural boundaries, semantic boundaries, temporal boundaries, and event boundaries to perform segmentation, merging, or aggregation on adjacent basic content units, generating at least one semantic evidence unit. Based on the source location address mapping of the basic content units covered by the semantic evidence unit, it generates the source location address mapping of the semantic evidence unit and attaches it to the semantic evidence unit, forming the smallest traceable knowledge unit, which serves as the address-based semantic evidence unit. By traversing all basic content units, multiple address-based semantic evidence units of the document to be processed are obtained for subsequent knowledge response package generation, evidence backreference, evidence continuity judgment, version comparison, or answer source write-back, etc. A semantic evidence unit includes at least one basic content unit.

[0098] Specifically, the segmentation boundaries of adjacent basic content units can be identified based on one or more of the following criteria: content boundaries, structural boundaries, semantic boundaries, temporal boundaries, and event boundaries. Adjacent basic content units are then segmented at these boundaries to obtain two semantic evidence units, with each semantic evidence unit corresponding to one basic content unit. To avoid over-segmentation of cross-page tables, cross-page lists, or the same event, adjacent basic content units or content fragments that meet the criteria can be merged or aggregated based on one or more of the following criteria to obtain semantic evidence units that include multiple basic content units. Content fragments are segments formed after merging or aggregating adjacent basic content units.

[0099] In one embodiment, the boundary score and merge score of adjacent basic content units can be determined, and based on one or more of the boundary score, merge score or rule determination of the basic content unit, the processing method for the adjacent basic content unit is determined: segmentation, merging or aggregation; the adjacent basic content units are processed based on the determined processing method to generate address-based semantic evidence units.

[0100] Specifically, the boundary score of adjacent basic content units can be determined based on one or more of the following: changes in title level, semantic similarity, page layout, time boundary, and block type. The boundary score is then used to determine whether a segmentation boundary has been formed. For example, if the boundary score is greater than or equal to the preset boundary score, no segmentation boundary is formed between adjacent basic content units, and segmentation is required. If the boundary score is less than the preset boundary score, a segmentation boundary is formed between adjacent basic content units.

[0101] The system determines the merge score for adjacent basic content units or content segments based on one or more of the following: title or table header similarity, column structure similarity, sequence number continuity, contextual continuity, pagination relationship, chain relationship of the same event, and relationship of the same version status. This merge score is used to determine whether adjacent basic content units or content segments should be merged or aggregated, avoiding excessive segmentation of cross-page tables, cross-page lists, or the same event. A linear weighted approach can be used to determine the boundary score and the merge score.

[0102] In this scheme, by determining the multidimensional boundary scores and merge scores of adjacent basic content units, and combining them with rule-based judgment, the scheme determines whether to perform segmentation, merging, or aggregation operations on the adjacent basic content units, thereby generating address-based semantic evidence units. When the boundary score is greater than or equal to a preset threshold, no segmentation boundary is formed; when it is less than the threshold, segmentation is performed. When the merge score meets the conditions, merging or aggregation is performed to avoid over-segmentation of cross-page tables, cross-page lists, or the same event, which significantly improves the semantic integrity and structural accuracy of the semantic evidence units.

[0103] In a specific embodiment, such as Figure 3 As shown, the generation process of address-based semantic evidence units specifically includes the following steps: S100: Input the document to be processed; S101: Extract document features from the document to be processed; S102: Select a target parsing path from multiple candidate parsing paths based on document features. The target parsing path includes at least one of structural parsing path, layout parsing path, image parsing path, and hybrid parsing path. S103: Based on the target parsing path, perform structural parsing, layout parsing, image parsing, or mixed parsing on the document to be processed to generate basic content units; S104: And analyze and extract the structural information corresponding to the basic content unit to obtain the layout structure information or logical structure information corresponding to the basic content unit. S105: Based on the layout or logical structure information corresponding to the basic content unit, reconstruct the reading order of the document to be processed to obtain the reading order number of the basic content unit. S106: Based on the reading sequence number, page layout information and / or logical structure information, generate a source location address mapping for the basic content unit that can point back to the source document location; S107: Perform boundary determination on basic content units based on content boundaries, structural boundaries, semantic boundaries, temporal boundaries, or event boundaries to determine the processing method for basic content units; S108: Based on a determined segmentation, merging, or aggregation method, perform one or more of segmentation, merging, or aggregation on the basic content unit to obtain multiple processed semantic evidence units; S109: Perform quality assessment and evidence hashing on semantic evidence units to obtain the quality score and evidence hash of the semantic evidence units; S110: Bind the semantic evidence unit with the corresponding source location address mapping, quality score, and evidence hash to obtain the address-based semantic evidence unit.

[0104] In this embodiment, by extracting document features from the document to be processed and adaptively selecting structural parsing, layout parsing, image parsing, or hybrid parsing paths, basic content units and their corresponding layout or logical structure information are generated. Then, reading order reconstruction and source location address mapping are performed. Based on content boundaries, structural boundaries, semantic boundaries, time boundaries, or event boundaries, the basic content units are segmented, merged, or aggregated. The resulting semantic evidence units are then subjected to quality evaluation and evidence hashing. Finally, the semantic evidence units are bound with source location address mapping, quality scores, and evidence hashes to form address-based semantic evidence units. This achieves traceability of evidence, completeness of semantics, adaptive parsing, and controllable quality, providing standardized and traceable basic knowledge units for the digital clone knowledge base.

[0105] In one embodiment, after generating address-based semantic evidence units, a quality score and evidence hash are calculated for each semantic evidence unit. Semantic evidence units with quality scores lower than a preset quality threshold are marked as low-confidence units or units to be reviewed, thereby triggering a rollback parsing (i.e., reprocessing the document content to be processed corresponding to the semantic evidence unit using other parsing paths) or manual review (i.e. prompting for manual review and correction). Alternatively, a downgrade process can be performed during subsequent calls, such as filtering low-confidence units during queries.

[0106] Specifically, the quality score of a semantic evidence unit can be determined based on one or more of the following: character recognition confidence, language coherence, structural integrity, source reliability, repetition, garbled text, address integrity, and parsing confidence. An evidence hash can be generated for the semantic evidence unit based on its corresponding document identifier, version identifier, source location address mapping, and evidence text. This evidence hash is a content fingerprint of the semantic evidence unit, primarily used for integrity verification, change detection, and version comparison, and is a key identifier ensuring the credibility of the knowledge base and the correctness of incremental updates.

[0107] S18: Perform one or more of the following on semantic evidence units: knowledge extraction, cross-document alignment, version organization, and index building, to form an evidence-binding knowledge base.

[0108] After obtaining multiple address-based semantic evidence units from the document to be processed, the knowledge extraction and alignment module performs knowledge extraction and binding on each semantic evidence unit to obtain a knowledge-bound semantic evidence unit. The temporal version chain construction includes entity extraction, relation extraction, event extraction, temporal normalization, and reference resolution; that is, the knowledge extraction and alignment module performs entity extraction, relation extraction, and event extraction on each semantic evidence unit to obtain a knowledge object for each semantic evidence unit. This knowledge object includes one or more of the following: entities, events, and relations between entities or events. During the extraction process, reference resolution is performed on the semantic evidence units to point pronouns (such as I, he, she, this project, etc.), abbreviations, aliases, etc., to specific entities.

[0109] Taking the document to be processed as a resume document as an example, the entities extracted should at least include people, organizations, projects, tasks, results, skills, and documents; the relationships extracted should at least include participation, responsibility, output, dependence, belonging, proof, and occurrence; the events extracted should at least include onboarding, project participation, delivery, awards, training, reporting, review, and resignation.

[0110] Then, the semantic evidence units are time-normalized to unify non-standard time expressions such as "last year," "second quarter," "as of now," and "late stage of the project" into a unified time interval, resulting in a standard time interval. The time precision and confidence level are then determined, yielding precision labels and time confidence levels. The precision label can be one of the following time precision levels: daily, weekly, monthly, quarterly, grade, or fuzzy. Finally, the knowledge object, standard time interval, precision label, and time confidence level are bound to the semantic evidence unit to form a semantic evidence unit carrying address mapping, extracted information, and time information.

[0111] The knowledge extraction and alignment module performs cross-document alignment on the same knowledge object based on the semantic evidence unit after knowledge binding, so as to merge different knowledge objects pointing to the same fact object in different documents into the same knowledge object, which is a node or relation edge in the fact graph; where the node of the fact graph is the same entity or event after cross-document alignment, and the relation edge of the fact graph is the same relation after cross-document alignment, representing the relationship between the nodes.

[0112] In one embodiment, for two candidate knowledge objects (such as entities, events, or relationships) in different documents, an alignment score is calculated for the pair of candidate knowledge objects. When the alignment score is greater than or equal to a preset alignment threshold, the two candidate knowledge objects are considered as mappings of the same knowledge object in different documents, and the mappings of candidate knowledge objects from different documents are merged into the same knowledge object so that they can be subsequently written into the fact graph to form nodes or relationship edges. Specifically, to avoid misalignment due to inconsistent knowledge object types, when the two candidate knowledge objects are of the same type (e.g., both are entities, events, or relationships) and the alignment score is greater than or equal to the preset alignment threshold, the mappings of the two candidate knowledge objects from different documents are merged into the same knowledge object so that they can be subsequently written into the fact graph to form nodes or relationship edges. When the two candidate knowledge objects are of different types or the alignment score is less than the alignment threshold, it indicates that the two candidate knowledge objects are not the same knowledge object, and they are not merged. This scheme achieves cross-document knowledge disambiguation and fusion, effectively avoiding knowledge duplication and conflict, and improving the accuracy and consistency of the fact graph.

[0113] The candidate knowledge objects are knowledge objects extracted from semantic evidence units, including at least one of entities, events, or relationships. The alignment threshold can range from 0.60 to 0.95.

[0114] Specifically, based on the semantic evidence units to which two candidate knowledge objects belong, the similarity in name, contextual semantic similarity, temporal overlap, and normalized role consistency and source consistency are calculated between the two candidate knowledge objects. Furthermore, based on these similarities, the alignment score is calculated for the two candidate knowledge objects from different documents. This alignment score can be expressed by the following formula: ; in, Indicates alignment score; Indicates the similarity between names or aliases; Indicates semantic similarity within the context; Indicates the degree of time overlap; This indicates role consistency; This indicates consistency in origin; This represents the pre-defined weight values.

[0115] After cross-document alignment, the aligned knowledge objects are written into the fact graph as nodes or relation edges, and the semantic evidence units covered by each node or relation edge and their source location address mappings are written into the corresponding node or relation edge as evidence binding mappings. That is, the fact graph includes a set of nodes, a set of relation edges, and evidence binding mappings.

[0116] The node set can represent entities (such as people, organizations, projects, skills, etc.) or events (such as onboarding, delivery, awards, etc.) in the semantic evidence unit; the relation edge set represents the relationship between nodes, such as participation, responsibility, output, occurrence, etc.; the evidence binding mapping set includes the evidence binding mapping associated with each node or relation edge, and each evidence binding mapping includes at least one semantic evidence unit, which is bound to a source location address mapping, quality score, and time information such as standard time interval, precision label, and time confidence.

[0117] For example, users upload three types of documents: resume documents, work logs, and project debriefing reports. The system generates document fingerprints for each document and determines the target parsing path based on document format and characteristics. For resume documents with clear style hierarchy and high text extraction rate, a structure parsing path can be selected; for work logs mainly using paragraphs and time expressions, a layout parsing path can be selected; and for project debriefing reports containing both scanned pages and raw text pages, a mixed parsing path can be selected. After parsing the basic content units and forming a layout tree, the system generates source location address mappings for the corresponding content based on information such as document identifier, version identifier, page number, normalized coordinates, reading order number, and table or cell path. Suppose the resume document states "responsible for the construction of the data platform for Project A", while the project debriefing report states "responsible for the delivery of the indicator platform for Project A". The system calculates an alignment score based on the similarity of the names, the semantic similarity of the context, the overlap of time, and the consistency of roles. When the alignment score exceeds a preset threshold, the two statements are mapped to the same knowledge object "Project A". Furthermore, an evidence-binding fact graph containing project nodes, responsibility nodes, result nodes, and time nodes is constructed to record the state evolution of "responsible for - output - delivery" in a time version chain.

[0118] The knowledge base organization and version indexing module, based on a fact graph, organizes the different states of each knowledge object across different documents to generate a temporal version chain for each knowledge object, i.e., a version organization structure for each knowledge object. The temporal version chain records the set of version states of the same knowledge object under different document versions, different effective times, and different supporting evidence. In other words, the temporal version chain records the version states of the same knowledge object at different times; each version state in the temporal version chain represents the specific description of that knowledge object in a certain time slice or a certain version of a document. For example, Project A's resume in 2023 describes it as "responsible for data platform construction," while the project review in 2024 describes it as "leading the delivery of the indicator platform." These two records represent two different version states of Project A.

[0119] In other words, the semantic evidence units associated with each knowledge object in the fact graph are sorted by time (e.g., standard time intervals). Semantic evidence units with similar time ranges, originating from the same document version, or with consistent expression are aggregated into a version state to generate version states of the same knowledge object in different time versions. The version states of different time versions are then organized into a time version chain for the knowledge object. This time version chain records the state of the same knowledge object in different document versions, realizing state evolution. If the same knowledge object has multiple different expressions (e.g., contradictory expressions) within the same document version, it is split into multiple conflicting versions under the same time version.

[0120] Within this framework, semantic evidence units with the same standard time interval and the same description are assigned a version status. The version status includes a set of semantic evidence units, a status identifier, an object identifier, an effective start time, an effective end time, a document version identifier, evidence support strength, source authority, source priority, a conflict marker (used to mark conflicting versions), conflict penalties, and evidence consistency.

[0121] The strength of evidence is calculated using the following formula: ; The authority of the source is calculated using the following formula: ; The conflict penalty and evidence consistency are calculated using the following formulas: ; ; in, Indicates the strength of the evidence; Indicates the authority of the source; Indicate consistency with the context; Indicates conflict penalty items; Indicates version status The set of semantic evidence units in; Represents semantic evidence unit The source weight is pre-calibrated based on the source type; This represents the quality score of the semantic evidence unit; A function for determining version conflicts between two semantic evidence units in the same version state; Represents ordered pairs in the set of semantic evidence units The total number; Indicates that ordered pairs satisfy The quantity.

[0122] The version status information can be used to calculate the version score of different version statuses of the knowledge object during the query, so as to select the best version status as the basis for the answer based on the version score in the time version chain of the knowledge object. Specifically, the version score of the version status can be calculated based on query time constraints, strength of evidence support, authority of the source, consistency of evidence and conflict penalty, so as to arbitrate multiple version statuses of the same knowledge object and select the target version status with the highest version score as the basis for the answer.

[0123] The version score for the version status is calculated using the following formula: ; in, Version rating indicating version status; Indicates the distance between the version status and the query time constraint; This represents the time decay coefficient, which can be dynamically updated based on feedback during operation. This indicates the pre-defined weights.

[0124] After generating the time-version chain, the knowledge base organization and version indexing module constructs indexes to establish the following retrieval indexes: object storage, full-text index, vector index, graph index, and version index. Specifically, object storage is used for semantic evidence units, knowledge objects, and document registration records; a full-text index is built for the document content to which the semantic evidence unit belongs; a vector index is built for the semantic vector of the semantic evidence unit; a graph index is built for nodes and edges in the fact graph; and a version index is built for the version status in the time-version chain.

[0125] The knowledge base organization and version indexing module associates and binds the fact graph, the semantic evidence units of each knowledge object in the fact graph, the version organization structure (i.e., the time version chain), and the corresponding index to form an evidence-bound knowledge base. That is, this evidence-bound knowledge base includes the set of semantic evidence units formed by all semantic evidence units, the knowledge objects associated with each semantic evidence unit, the version organization structure of each knowledge object, and the retrieval index.

[0126] S19: Upon receiving the training request, the knowledge layer and style layer of the digital clone are trained separately.

[0127] That is, after building or updating the evidence binding knowledge base, if a digital clone adjustment request is received, such as a training request, parameter calibration request, configuration request (including driver configuration and prompt configuration), and invocation strategy optimization request, the system responds to the corresponding request and performs one or more of the following on the digital clone based on the evidence binding knowledge base: training, parameter calibration, configuration, or invocation strategy optimization; wherein, during the training process, the knowledge layer and style layer of the digital clone are trained separately.

[0128] Upon receiving a training request, the knowledge base organization and version indexing module can construct training samples based on multiple semantic evidence units in the evidence-binding knowledge base, generating a training sample set for training the digital avatar. The hierarchical training module performs training on the specified digital avatar, separating the knowledge layer and style layer of the digital avatar during training to enable question-and-answer interaction based on the trained digital avatar. The training of the digital avatar aims to prioritize the knowledge layer in fact selection to ensure correct and traceable answers, and to follow the style layer in expression to mimic the user's personal expression habits. The decoupling of the knowledge layer and style layer training reduces the interference of style transfer on factual content.

[0129] In one embodiment, when simultaneously performing knowledge layer training and style layer training, the knowledge layer and style layer are trained separately through one or more of the following methods: orthogonal constraints, gating isolation, mutual information minimization, projection isolation, parameter isolation, cue isolation, sample isolation, or output verification isolation. This reduces the coupling between knowledge representation and style representation, allowing the digital clone to prioritize the knowledge layer in fact selection and follow the style layer in expression, thereby reducing the interference of style transfer on factual content.

[0130] In one embodiment, the knowledge layer training is used to optimize one or more of the following: evidence matching, recall ranking, version consistency, reference location, evidence coverage, permission verification, and rejection judgment; the style layer training is used to optimize one or more of the following: terminology preference, sentence structure preference, length preference, narrative perspective, output structure, and grouping expression.

[0131] In one embodiment, during training, multiple semantic evidence units in the training sample set are first divided into positive and negative evidence sets to form a positive evidence set and a negative evidence set. The positive evidence set contains multiple semantic evidence units that can support the standard answer, i.e., multiple positive evidence samples. The negative evidence set contains multiple semantic evidence units that are irrelevant to the standard answer, contradictory, or of low quality, i.e., multiple negative evidence samples. Then, structured noise sampling is performed on the positive evidence set to obtain multiple sampled evidence to generate a perturbed evidence set. Knowledge layer training is then performed based on these multiple sampled samples to enhance the model's robustness. The structured noise sampling includes at least one or more of the following: evidence deletion, similar evidence replacement, evidence masking, and cross-version replacement.

[0132] In this process, style features are extracted from users' historical self-descriptions, summaries, replies, and edit records to form a set of style features for style layer training.

[0133] In one embodiment, knowledge layer training may employ one or more loss functions selected from ranking loss, reference location loss, temporal consistency loss, evidence coverage loss, and robustness loss as training objectives. Style layer training may employ one or more of style classification, expression habit consistency, length control, and narrative perspective control as training objectives.

[0134] After constructing the evidence binding knowledge base, during the operation of the digital clone, it receives instructions to add, update, or delete documents, or permission changes. Based on these instructions, it incrementally updates the existing evidence binding knowledge base, forming a new version. The digital clone is then trained based on this new version to achieve timely updates and iterations of the knowledge base. Figure 2 As shown, incremental updates specifically include the following steps: S20: During the operation of the digital clone, receive instructions to add or update documents, delete documents, or change permissions.

[0135] S21: Upon receiving instructions to add a document, update a document, delete a document, or change permissions, perform a hierarchical fingerprint comparison of the corresponding document to identify the set of changes.

[0136] S22: Perform differential reconstruction, update or invalidation processing on one or more of the semantic evidence units, knowledge objects, version states, permission tags or index items corresponding to the change set to form a new version of the evidence binding knowledge base, and perform step S19 and subsequent steps based on the new evidence binding knowledge base to realize digital clone training, invocation, parameter calibration and online update.

[0137] Upon receiving a new document, an updated document, or a deletion instruction or permission change information, the incremental update and evaluation module generates a hierarchical fingerprint of the changed document or permission change information to identify the changed object based on the hierarchical fingerprint of the changed document or permission change information. The changed object includes one or more of the following: basic content unit, semantic evidence unit, knowledge object (i.e., node or relation edge of fact graph), version status, permission tag or index item. After identifying all changed objects, they are combined into a change set.

[0138] The incremental update and evaluation module, based on the corresponding scenario, performs differential reconstruction, update, deletion, or invalidation processing on one or more of the semantic evidence units, knowledge objects, version states, permission tags, or index items corresponding to the change set. The differential reconstruction (i.e., incremental reconstruction) process can be referred to in steps S12 to S18 to identify the changed objects corresponding to the newly added or updated documents, perform differential reconstruction, thereby generating newly added or updated address-based semantic evidence units, and then perform knowledge extraction, cross-document alignment, version organization, and index construction to update and obtain a new version of the evidence-binding knowledge base.

[0139] In this embodiment, during the construction of the evidence binding knowledge base and the operation of the digital clone, by receiving new documents, updated documents, deletion instructions, or permission change information, hierarchical fingerprint comparison is performed on the changed objects to identify the change set. Differential reconstruction, updating, or invalidation processing is performed on the semantic evidence units, knowledge objects, version status, permission tags, or index items corresponding to the change set, thereby forming a new version of the evidence binding knowledge base. Based on the new version of the knowledge base, the digital clone is trained, invoked, parameter calibrated, and updated online. This achieves timely and low-cost iterative updates of the knowledge base, significantly improving the real-time performance and maintainability of the digital clone knowledge base in scenarios with frequent document revisions and dynamic permission changes.

[0140] In one embodiment, after constructing or updating the evidence binding knowledge base and training the digital clone, it is further used to execute corresponding query or invocation tasks based on the trained digital clone, thereby realizing the application of the digital clone, such as... Figure 2 As shown, the specific steps include the following: S23: When a query or call request is received during the operation of the digital clone, query constraint information is generated based on the query or call request. The query constraint information includes one or more of the following: time constraint, permission constraint, evidence constraint, style constraint, context budget constraint, and latency budget constraint.

[0141] When a query or call request is received during the operation of the digital clone, the query orchestration module generates query constraint information based on one or more of the following constraints: time constraint, permission constraint, evidence constraint, style constraint, context budget constraint, and latency budget constraint. This information is then used for subsequent retrieval and processing of semantic evidence units in the knowledge base.

[0142] S24: Based on the query constraint information, perform candidate filtering and recall sorting on the semantic evidence units in the evidence binding knowledge base to obtain the recall evidence set.

[0143] Based on the query constraint information, the candidate pre-filtering and controlled loading module performs candidate filtering, recall sorting, controlled loading, and permission pruning on multiple semantic evidence units in the evidence binding knowledge base, thereby obtaining one or more loaded semantic evidence units to form a loaded evidence set.

[0144] Specifically, the time-version chain records the set of version states of the same knowledge object under different document versions, different effective times, and different supporting evidence. Multiple version states in the set can be arbitrated based on query time constraints, evidence support strength, source authority, evidence consistency, and conflict penalties to select the target version state as the basis for the answer; the arbitration process is described above and will not be repeated here. Then, based on the query constraint information, candidate filtering, recall ranking, controlled loading, and permission pruning are performed on multiple semantic evidence units corresponding to the target version state to obtain one or more loaded evidences, forming a loaded evidence set. This achieves version-aware knowledge retrieval and loading, significantly improving the accuracy, consistency, and traceability of digital avatar question answering in time-sensitive and version-evolving scenarios.

[0145] In one embodiment, prior to hybrid recall and controlled loading, candidate semantic evidence units are pre-filtered. This pre-filtering stage includes: pre-filtering multiple candidate semantic evidence units based on query time constraints (e.g., requiring semantic evidence units to be effective within a specific time window), permission constraints (e.g., the user's permission level and tags must not be lower than the permissions required by the semantic evidence unit), and a preset evidence quality threshold (i.e., evidence constraints). Semantic evidence units that do not meet the time conditions, permission conditions, or quality scores below the threshold are removed, resulting in multiple candidate semantic evidence units that meet the conditions, forming a candidate set. This candidate set is used for subsequent hybrid recall, sorting, controlled loading, and permission pruning to obtain the loaded evidence set.

[0146] Specifically, regarding the time condition, the standard time interval corresponding to the semantic evidence unit is matched with the query time window; when the query does not contain time constraints, or the standard time interval of the semantic evidence unit and the query time window meet a preset overlap condition, the semantic evidence unit is determined to meet the time condition. Regarding the permission condition, the judgment is made based on the permission level, permission tag, and permission tag of the semantic evidence unit; when the query subject has the permission to access the evidence unit, the semantic evidence unit is determined to meet the permission condition. Regarding the quality condition, the judgment is made based on one or more of the following: quality score, address integrity, resolution confidence, and verification mark; semantic evidence units with a quality score less than a preset quality threshold or with incomplete addresses are not included in the candidate set, or are used as low-confidence supplementary evidence.

[0147] Then, a recall score is calculated for each semantic evidence unit in the candidate set, and a hybrid recall is performed on the semantic evidence units in the candidate set based on the recall score, that is, multi-way recall of candidate semantic evidence units based on the query statement. The recalled candidate semantic evidence units are then sorted to obtain a sorted recall evidence set. The recall score is calculated based on one or more of the following: semantic similarity between the semantic evidence unit and the query statement, keyword matching degree, graph correlation degree, time matching degree, source weight, permission satisfaction degree, evidence completeness, and conflict penalty. For example, the weighted sum of the indicators of each dimension can be used to obtain a hybrid recall score for the semantic evidence unit, and the semantic evidence units in the candidate set can be sorted in descending order accordingly to obtain a sorted recall evidence set.

[0148] Among them, semantic similarity is used to measure the semantic association between the query statement and the semantic evidence text; keyword matching degree is used to measure the literal matching between keywords in the query statement and the semantic evidence text; graph association degree is used to measure the association between entities, relations or events involved in the query statement and nodes and edges in the fact graph; time matching degree is used to measure the degree of consistency between the evidence time and the query time; source weight is used to reflect the credibility of the document source or version source; evidence completeness is used to reflect whether the evidence address, context and version information are complete; conflict penalty is used to reduce the ranking priority of semantic evidence units with version conflict or fact conflict.

[0149] In this scheme, before hybrid recall and controlled loading, candidate semantic evidence units are filtered based on query time constraints, permission constraints, and preset evidence quality thresholds. Evidence units that do not meet the time conditions, permission conditions, or have quality scores below the thresholds are eliminated, forming a candidate set. Subsequently, for each semantic evidence unit in the candidate set, a recall score is calculated and sorted based on one or more of the following: semantic similarity, keyword matching degree, graph correlation degree, time matching degree, source weight, permission satisfaction degree, evidence completeness, and conflict penalty. This results in a sorted recall evidence set. Thus, this scheme achieves accurate pre-filtering of evidence and multi-dimensional hybrid recall, significantly improving the efficiency of subsequent controlled loading and the quality of the loaded evidence set.

[0150] S25: Perform controlled loading and permission trimming on the semantic evidence units in the recalled evidence set to obtain the loaded evidence set.

[0151] Controlled loading does not simply select a fixed number of top-ranked pieces of evidence. Instead, it dynamically selects the set of evidence to be loaded under constraints of a context budget (the total length of evidence allowed to be loaded) and a latency budget (the allowed processing time). Controlled loading includes: applying loading constraints to the recalled evidence set; selecting candidate semantic evidence units from the recalled evidence set based on the loading constraint results and recall scores; and jointly loading the selected candidate semantic evidence units based on evidence continuity rules to obtain the loaded evidence set. Loading constraints include one or more of the following: context budget, latency budget, permission constraints, and evidence quantity constraints. Evidence continuity rules include one or more continuity relationships such as the same source document, adjacent address paths, the same table, the same list, the same chapter, the same event chain, or the same version state.

[0152] Specifically, in the loading evidence set, candidate semantic evidence units are traversed in sorted order. For each semantic evidence unit, its length, expected processing delay, and permission status are checked. Based on these parameters, the semantic evidence units are loaded one by one according to the recall score, forming the loading evidence set. If the length, expected processing delay, and permission status of a semantic evidence unit meet the context budget, delay budget, and permission constraints, it is loaded into the loading evidence set. During this process, it is checked whether there is a continuity relationship between the current semantic evidence unit and semantic evidence units in the loading evidence set, such as whether they are from the same source document, adjacent address paths, the same table, the same list, the same chapter, the same event chain, or the same version status. If a continuity relationship exists, the joint loading priority of the semantic evidence unit is increased, allowing it to be loaded into the loading evidence set under the same conditions. When the number of semantic evidence units in the loading evidence set exceeds the evidence quantity constraint or the delay is about to exceed the limit, loading stops, thus obtaining a loading evidence set that meets the loading constraints and has high relevance and continuity.

[0153] During the controlled loading process, when the length of the semantic evidence unit and the expected processing delay meet the context budget and delay budget, the permission and security control module can perform permission pruning on the semantic evidence unit based on permission constraints to hide, obscure, or generalize knowledge fragments that exceed the permission constraints, thereby obtaining semantic evidence units that meet the permission constraints. The semantic evidence units that meet the permission constraints are then loaded to obtain the loaded evidence set.

[0154] In this solution, a controlled loading mechanism is used to dynamically select and load candidate semantic evidence units to form a loading evidence set under loading constraints such as context budget, latency budget, permission constraints, and evidence quantity constraints, combined with recall scoring and evidence continuity rules (including one or more continuous relationships among the same source document, adjacent address paths, the same table, the same list, the same chapter, the same event chain, or the same version state). This avoids the problems of continuous information breakage, incomplete context, and budget overrun caused by simply truncating the top-ranked evidence, and significantly improves the evidence relevance, semantic continuity, and response reliability of digital avatar question answering under limited resources.

[0155] S26: A knowledge response package for generating a digital clone based on a set of loaded evidence.

[0156] After generating the loaded evidence set, the knowledge response package generation module generates digital clone response information based on the semantic evidence units in the loaded evidence set, so as to generate a knowledge response package in response to query or call requests; the knowledge response package includes the digital clone's answer content and response control information.

[0157] The response control information includes one or more of the following: evidence address set, evidence identifier, permission mask, version snapshot, style tag, response confidence level, degradation status, session consistency identifier, group sequence number, and pending review marker. When the response control information includes an evidence address set or an evidence identifier, the evidence address set or evidence identifier is associated with at least one semantic evidence unit in the loaded evidence set.

[0158] Specifically, the knowledge response package generation module generates response content for queries or invocation requests based on semantic evidence units in the loaded evidence set, and generates response control information based on the response content and the loaded evidence set. Specifically, it determines the evidence coverage status based on the degree of matching between the response content and the loaded evidence set, the source referential status based on the integrity of the evidence address in the loaded evidence set, the permission status based on the permission pruning result, and determines the response confidence level based on one or more of the following: evidence quality (i.e., the quality score of the semantic evidence unit), evidence completeness, response consistency, proportion of uncovered content, and permission verification results. Simultaneously, it generates a snapshot of the target version state used to generate the loaded evidence set, resulting in a version snapshot.

[0159] In one embodiment, the permission and security control module performs permission pruning on semantic evidence units, candidate fragments, or response fragments before or during candidate filtering, controlled loading, answer generation, or response output. This permission pruning supports at least one or more granularities: document-level, chapter-level, paragraph-level, table-level, field-level, evidence-level, or response group-level. By performing permission pruning at at least one or more granularities at the document-level, chapter-level, paragraph-level, table-level, field-level, evidence-level, or response group-level on semantic evidence units, candidate fragments, or response fragments, fine-grained, multi-layered control over knowledge access permissions is achieved. This effectively prevents unauthorized output of sensitive information and significantly improves the security and compliance of the digital avatar knowledge base in multi-party sharing and hierarchical authorization scenarios.

[0160] Specifically, for semantic evidence units, candidate segments within semantic evidence units, or response segments within answer content, a visibility mask (i.e., a permission mask) is calculated based on permission constraints. Semantic evidence units or segments with a visibility mask equal to 0 are hidden, obfuscated, or generalized to a range, while semantic evidence units or segments with a visibility mask equal to 1 are displayed, thus obtaining semantic evidence units that satisfy the permission constraints. The visibility mask can be represented by the following formula: ; in, Indicates a visibility mask; Indicates the permission level of the query subject; Indicates the permission level required by a semantic evidence unit or fragment; A set of permission tags representing the query subject; A set of permission tags representing semantic evidence units or fragments; Indicates an indicator function; This represents the empty set.

[0161] S27: Based on the evidence coverage status, permission status, version status, or confidence status of the knowledge response package, output one of the following: normal response, downgraded response, rejection prompt, or pending review prompt.

[0162] After generating the knowledge response package, the system controls the digital avatar to respond to query or call requests based on the knowledge response package. The digital avatar's dialogue interface module drives the digital avatar to output one of the following: a normal response, a downgraded response, a rejection message, or a message awaiting review. Specifically, when the response confidence and evidence coverage status meet the normal output conditions, the system controls the digital avatar to output a normal response.

[0163] In a specific embodiment, such as Figure 4 As shown, the process of generating a knowledge response package based on the evidence-binding knowledge base and executing the corresponding response output upon receiving a user query or call request includes the following steps: S200: Receives user queries or call requests; S201: Parse the query constraints carried in the query or call request to obtain one or more of the following constraints: time constraint, permission constraint, evidence constraint, style constraint, context budget constraint, and latency budget constraint; S202: Summarize the parsed query constraints into a query constraint object, i.e., query constraint information; S203: Filter candidate semantic evidence units in the evidence binding knowledge base based on query constraint information; S204: Perform multi-path candidate recall on the filtered candidate semantic evidence units; S205: Perform a unified sorting of the recalled candidate semantic evidence units; S206: Perform controlled loading on the sorted candidate semantic evidence units to load the candidate semantic evidence units that satisfy the query constraints into the loading evidence set; S207: During the loading process, evidence loading is performed synchronously to prioritize loading semantic evidence units that have a continuous relationship with candidate semantic evidence units in the loaded evidence set into the loaded evidence set; S208: Perform permission trimming on the loaded semantic evidence units to obtain a set of loaded evidence that satisfies the query constraint information; S209: Generate a knowledge response package based on the loaded evidence set; S210: Determine the evidence coverage status, permission status, version status, and confidence status of the knowledge response package; S211: Based on the judgment result, output the normal output, the summary downgraded output, or the rejection prompt output; S212: The digital clone calls the corresponding output to respond.

[0164] In this embodiment, by receiving user queries or invocation requests and parsing multi-dimensional constraints such as time, permissions, evidence, style, context budget, and time delay, and summarizing them into query constraint objects, the semantic evidence units in the evidence binding knowledge base are sequentially subjected to candidate filtering, multi-way candidate recall, unified sorting, controlled loading, and continuous evidence loading. During the loading process, permission pruning is performed simultaneously to obtain a set of loaded evidence that meets the query constraints. Then, a knowledge response package is generated, and its evidence coverage status, permission status, version status, and confidence status are judged. Finally, based on the judgment result, a normal response, a summary downgrade response, or a rejection prompt response is output for the digital avatar to call and respond. This achieves accurate screening, continuous loading, and permission security control of evidence under resource budget constraints, significantly improving the evidence traceability, response controllability, and output security of the digital avatar question answering system.

[0165] In one embodiment, when any of the following conditions are met: response confidence is lower than a preset threshold, evidence coverage is lower than a preset threshold, permission verification fails, or version conflict cannot be arbitrated, a downgraded response, a rejection message, or a pending review message is output. The downgraded response includes one or more of the following: summary downgrade output, insufficient evidence message, insufficient permission message, and version conflict message.

[0166] Specifically, when the response confidence level is below a preset threshold, the evidence coverage is below a preset threshold, permissions are insufficient, there is a version conflict, or the latency budget is insufficient, but a general answer can still be generated, a downgraded response is output. When key facts lack supporting evidence, permission verification fails, or version conflicts prevent arbitration, a refusal prompt or control prompt is output. This solution can avoid outputting unreliable or unauthorized content in scenarios where high-quality, credible, and compliant answers cannot be generated, significantly improving the output security, compliance, and user experience of the digital avatar system.

[0167] In one embodiment, the knowledge response package can be divided into multiple response groups, that is, the knowledge response package generation module splits the generated knowledge response package into multiple response groups output in chronological order; each response group includes at least one answer fragment of the answer content and response control sub-information. The response control sub-information includes one or more of the following: evidence address subset, evidence identifier subset, permission mask subset, version snapshot, style tag, time sequence number, session consistency identifier, group confidence, and pending review mark.

[0168] This solution splits the knowledge response package into multiple response groups that are output in chronological order. Each group carries response fragments and corresponding response control sub-information such as evidence address subsets, evidence identifier subsets, permission mask subsets, version snapshots, style tags, time sequence numbers, session consistency identifiers, group confidence levels, and pending verification marks. This enables the digital avatar dialogue client to perform real-time, segmented evidence tracing, permission verification, and version consistency checks on streaming responses. It effectively avoids the problem of broadcasting before verification caused by the entire output, improves the response speed and user experience of real-time dialogue, and ensures the consistency between the front-end version and the knowledge version and the security and controllability of the output content in the same round of dialogue.

[0169] In one embodiment, a session consistency identifier is generated before the start of each round of dialogue or before the output of response groups to support real-time dialogue invocation by the digital avatar dialogue initiator. This session consistency identifier is determined based on one or more of the following: front-end avatar version, knowledge version, style version, permission version, model version, and round sequence number, and is written into at least one response group in the same round of dialogue. This scheme ensures that response groups output in the same round of dialogue carry the same session consistency identifier, reducing the risk of inconsistencies in the front-end avatar version, knowledge version, and style version, and ensuring version alignment and consistent output of the digital avatar during real-time dialogue invocation.

[0170] In one embodiment, a knowledge response packet or response group is output to the digital avatar's dialogue invocation client, knowledge Q&A invocation client, or knowledge service interface. The digital avatar's dialogue invocation client includes a front-end invocation client constructed based on one or more of the following: reference image, reference video, 3D avatar parameters, visualized character parameters, voice parameters, action parameters, or text interaction parameters. This dialogue invocation client is used to perform real-time dialogue invocation based on one or more of the response group's time sequence number, session consistency identifier, permission mask, version snapshot, or style tag.

[0171] The front-end of this dialogue invocation client includes an image parameter module, a broadcast module, and an action or expression driving module. The image parameter module can be constructed from one or more of the following: reference images, reference videos, 3D image parameters, or visualized character parameters. The broadcast module and the action or expression driving module perform real-time dialogue invocation based on one or more of the following: response group time sequence number, session consistency identifier, permission mask, and style tags. This solution achieves unified and collaborative invocation of the digital avatar front-end image, broadcast content, knowledge evidence, version status, permission boundaries, and style tags, significantly improving the coherence, security, and user experience of real-time dialogue.

[0172] In one embodiment, before outputting a response group or knowledge response packet, the system determines whether the response group or knowledge response packet is allowed to be submitted to the digital avatar's dialogue caller based on the group's confidence level, permission verification result, evidence coverage status, and remaining latency budget. If submission is allowed, the response group or knowledge response packet is sent to the digital avatar's dialogue caller; if submission is not allowed, the response group or knowledge response packet is discarded. Alternatively, the response group can be delayed in output, replaced with a downgrade prompt, or merged into a subsequent response group. This achieves pre-verification of output quality and security during real-time dialogue, preventing content with low confidence, insufficient permissions, or incomplete evidence coverage from being directly broadcast, significantly improving the credibility and security of the digital avatar's output.

[0173] In a specific embodiment, taking a query request as an example, the specific process of loading the evidence set generated in response to the query request, and executing the response output based on the knowledge response package, is as follows: Figure 5 As shown, the specific steps include the following: S001: Receive query requests sent by users through user terminals; S002: In response to the query request, the semantic evidence units in the evidence binding knowledge base are subjected to candidate filtering, recall sorting, controlled loading and permission pruning, and then a loaded evidence set is generated. S003: Generate response text, evidence address set, permission mask, version snapshot, style tags, and response confidence information based on the loaded evidence set; S004: Organize the generated response text, evidence address set, permission mask, version snapshot, style tags, and response confidence information into a knowledge response package; S005: Perform response verification (i.e. evidence coverage and confidence judgment) based on the knowledge response package, and output one of the following based on the verification result: normal output, summary downgrade output, or rejection prompt output; S006: Send the normal output, summary downgraded output, or rejection prompt output to the digital clone's dialog call terminal for query response.

[0174] For example, the query statement corresponding to the query request is: Please summarize the deliverables I was responsible for in the first half of 2024 and provide the source.

[0175] The system first parses the query time window as [2024-01-01, 2024-06-30], and identifies "Give source" as the evidence constraint for query constraints and "Please summarize" as the answer constraint for generating an answer. The system then generates a candidate set from the vector index, full-text index, graph index, and version index corresponding to the evidence-binding knowledge base. The recall scores of the five candidate semantic evidence units are set to 0.93, 0.88, 0.67, 0.91, and 0.52, respectively. The first and fourth candidate semantic evidence units come from adjacent paragraphs of the same project document, resulting in high scores for evidence continuity; the second candidate semantic evidence unit is relevant but comes from a historical version; and the fifth candidate semantic evidence unit does not meet the permission requirements. The system completes candidate pre-filtering and controlled loading based on the query time window, permission tags, evidence quality, recall score, evidence continuity, context budget, and latency budget. In scenarios where context budget and latency budget are limited, the system prioritizes jointly loading evidence units from adjacent paragraphs of the same project document and uses historical version evidence as supplementary evidence or conflict arbitration evidence; the budget can be dynamically configured according to the deployment environment, caller type, permission level and real-time requirements.

[0176] Subsequently, the system generates a knowledge response package, which includes: "Summary of deliverables handled in the first half of 2024". The evidence set includes the corresponding page number, paragraph path, and cell path of the semantic evidence unit, as well as the current permission mask, time version snapshot, and user's summary style tags. If the system finds that a sentence to be output is not fully covered by the loaded evidence set, it lowers the confidence of the response segment containing that sentence and replaces the sentence with an insufficient evidence prompt, a downgraded summary, or a rejection prompt.

[0177] In this embodiment, by receiving a user query request, candidate filtering, recall sorting, controlled loading, and permission pruning are performed on the semantic evidence units in the evidence binding knowledge base to generate a loaded evidence set. Based on this set, information such as answer text, evidence address set, permission mask, version snapshot, style tag, and response confidence is generated and organized into a knowledge response package. Then, response verification (evidence coverage and confidence judgment) is performed. Based on the verification result, normal output, summary downgrade output, or rejection prompt output is selected and sent to the dialogue call terminal of the digital clone for query response. This realizes a complete closed loop from query to loading, organization, verification, and hierarchical output, which significantly improves the evidence traceability, security controllability, and response reliability of the digital clone question answering.

[0178] In one embodiment, after constructing or updating the evidence binding knowledge base and training the digital clone, parameter calibration, configuration, and invocation strategy configuration can be performed on the digital clone before or during its operation, as well as online updates of parameters, configurations, or invocation strategies, to improve the performance of the digital clone and the system. For example, Figure 2 As shown, the parameter calibration and online update process can be found in the following steps: S28: During the operation of the digital clone, the parameters of the digital clone are calibrated and updated online to drive the calibrated and updated digital clone to perform the corresponding tasks.

[0179] That is, after the evidence binding knowledge base is built or updated, during the operation of the digital clone, when a request for parameter calibration, configuration or call strategy optimization is received from the digital clone, the hierarchical training module responds to the corresponding request and performs parameter calibration, configuration or call strategy optimization on the digital clone based on the built or updated evidence binding knowledge base.

[0180] In one embodiment, upon receiving a parameter calibration request, the hierarchical training module also responds to the parameter calibration request by performing offline or online parameter calibration on the digital clone; wherein, the calibration parameters include parameters that affect the calculation results of each stage, such as parsing path weight, boundary threshold, version arbitration weight, controlled loading weight, and response threshold.

[0181] During the offline parameter calibration process, the hierarchical training module tests the operation of the digital clone based on the validation set, historical query logs, manual correction records, or user feedback records, and detects the target metrics. When the digital clone is online, the target metrics are calculated based on the actual operating data of the digital clone. After obtaining the target metrics, the calibration parameters are calibrated offline and progressively updated online with the goal of reducing the target metrics. The target metrics include one or more of the following: address return error, time version selection error rate (i.e., version state selection error rate), permission pruning error rate (unauthorized access or false masking), response latency exceeding the limit rate, and the proportion of answers that do not cover evidence. The progressive update can adopt one or more of the following methods: fixed step size update, sliding window update, scenario-based parameter update, or update after manual review.

[0182] That is, based on one or more of the following factors during the testing or operation of the digital clone: ​​address back pointer error, time version selection error rate, permission pruning error rate, response latency exceeding the limit rate, and the proportion of evidence not covered in the response, one or more of the following parameters are calibrated and updated: parsing weight, boundary threshold, version arbitration weight, controlled loading weight, and response threshold, so as to perform relevant tasks based on the calibrated or updated digital clone.

[0183] The calibration parameters described above are merely illustrative; in other embodiments, they may include other parameters. Furthermore, the calibration parameters and their value ranges can be dynamically configured based on document type, business scenario, permission level, tenant configuration, historical query logs, manual review results, or online feedback records. Different parameter groups and their value ranges can be used for different document sources, different real-time requirements, or different permission levels, and updated through validation set evaluation, manual correction records, or online feedback records.

[0184] In this embodiment, based on the constructed or updated evidence-binding knowledge base, the digital clone is calibrated and progressively updated offline or online. The calibration parameters and their value ranges can be dynamically configured according to document type, business scenario, permission level, tenant configuration, historical query logs, manual review results, or online feedback records. Different document sources, real-time requirements, or permission levels can use different parameter groups and be updated through validation set evaluation, manual correction records, or online feedback records. This achieves adaptive optimization and continuous performance improvement of the digital clone knowledge base system, significantly enhancing the system's robustness, maintainability, and scenario adaptability.

[0185] In one embodiment, when performing incremental reconstruction on the evidence-binding knowledge base, the incremental reconstruction process specifically includes: generating hierarchical fingerprints at the paragraph, table, cell, chapter, page, event, field, or address path levels for changed documents or permission change information. These fingerprints can be generated through hash operations based on the content text, structural path, version identifier, and permission tags of the corresponding level. Based on the generated hierarchical fingerprints, a change set is identified, and one or more of the semantic evidence units, knowledge objects, version states, permission tags, and index items corresponding to the change set are reconstructed, updated, or invalidated. When an object in the change set has associated objects with continuity or dependency, a linked reconstruction is performed on the objects in the change set and the associated objects. Associated objects include adjacent content of the same document, and associated units of the same knowledge object, the same event chain, or the same version state. By performing linked reconstruction on the objects in the change set and the associated objects, fine-grained and high-precision incremental updates are achieved, effectively avoiding knowledge breaks or inconsistencies caused by isolated updates, and significantly improving the integrity and consistency of the digital clone knowledge base in dynamic change scenarios.

[0186] Specifically, when the document to be processed is a changed document, that is, when a new document or an updated document of an existing document is received, the system does not perform a full reconstruction of the entire knowledge base. Instead, it generates a hierarchical fingerprint of the content corresponding to the changed document, compares the hierarchical fingerprint of the new version document with the hierarchical fingerprint of the old version, identifies the changed objects based on the hierarchical fingerprint, and obtains a set of changes. Based on one or more of the semantic evidence units, knowledge objects, version status, permission tags or index items corresponding to the set of changes, differential reconstruction is performed to reconstruct, update or invalidate the corresponding objects, forming a new version of the evidence binding knowledge base.

[0187] Hierarchical fingerprints can be generated based on one or more of the following: structural unit text, structural path, time attribute, permission tag, version identifier, address path, and content type. Structural unit text includes one or more hierarchical levels such as paragraph level, table level, cell level, chapter level, page level, event level, field level, or address path level.

[0188] Correspondingly, when permission change information is received, the object whose permission has changed in the evidence binding knowledge base is identified based on the permission change information, and is used as the changed object to form a change set; the object whose permission has changed includes one or more of the following: semantic evidence unit, knowledge object (i.e., node or relation edge of fact graph), version status, permission tag or index item; the permission information of the corresponding changed object (such as permission level or tag) is changed to update the knowledge base and form a new version of the evidence binding knowledge base.

[0189] Upon receiving a deletion instruction, the system identifies objects to be deleted in the evidence-binding knowledge base based on the deletion instruction, such as semantic evidence units, knowledge nodes, knowledge edges, version status, permission tags, or index items, and treats them as changed objects to form a change set. The system then deletes or invalidates the semantic evidence units, knowledge nodes, knowledge edges, version status, permission tags, or index items in the evidence-binding knowledge base that correspond to the change set, thereby updating the knowledge base and forming a new version of the evidence-binding knowledge base.

[0190] In addition, when performing differential reconstruction, update, deletion or invalidation processing, for objects in the evidence-binding knowledge base that have adjacent content relationships, the same event chain relationships, the same knowledge object relationships or the same version state relationships with the change set, linkage reconstruction is performed to maintain evidence continuity and version consistency, which can reduce the reconstruction scope when adding or updating documents and improve continuous update efficiency.

[0191] In a specific embodiment, taking the receipt of a new or updated document as an example, the process of performing an incremental update on the evidence binding knowledge base is as follows: Figure 6 As shown, the specific steps include the following: S300: Users upload new versions of existing topic documents to the system through user terminals, such as adding or updating documents; S301: The corresponding module of the system receives the new version document and generates a hierarchical fingerprint to save the hierarchical fingerprint to the storage cluster; S302: The system's incremental update and evaluation module requests the hierarchical fingerprint of the old version of the document from the storage cluster; S303: The storage cluster returns the corresponding fingerprint result; S304: The system's incremental update and evaluation module compares the hierarchical fingerprints of the old and new versions of the document to determine the affected units (i.e. changed objects) in the evidence binding knowledge base, and requests the storage cluster to return the change set corresponding to the changed object; S305: The system's incremental update and evaluation module reconstructs the affected semantic evidence units based on this set of changes and returns the reconstruction results; S306: After successful reconstruction, the system's incremental update and evaluation module updates the corresponding index, fact graph, and version status in the evidence binding knowledge base based on the reconstruction results, so as to form a new system index, fact graph, and time version chain (referred to as index graph version chain), and returns the update completion status; S307: The system's document access and version registration module returns an update completion status to the user terminal so that the user is aware of the update status of the evidence binding knowledge base.

[0192] If the changed set only includes two structural units: the "Responsibility Description Paragraph" and the "Quantitative Results Table," then only the corresponding semantic evidence units, fact graph edges, and index items are reconstructed, without reconstructing other unchanged content. When another user queries "What are the main results of Project A?", if their permission level is insufficient to view the quantitative values, the relevant fields are anonymized according to the permission pruning rules, and only a general description such as "Completed the delivery of a certain indicator platform and formed a result accumulation" is returned; if the permission is sufficient, a complete result description and evidence address are returned.

[0193] In this embodiment, taking the receipt of new or updated documents as an example, the system generates a hierarchical fingerprint and saves it to the storage cluster after the user uploads a new version of the document. After requesting the hierarchical fingerprint of the old version of the document from the storage cluster and comparing it, the affected units and change sets are determined. Based on the change sets, the affected semantic evidence units are reconstructed, and then the corresponding indexes, fact graphs and version states in the evidence binding knowledge base are updated to form new indexes, graphs and version chains. Finally, the update completion status is returned to the user, thereby realizing the incremental update of new version documents of existing topic documents, significantly reducing the update cost of the knowledge base, and improving the real-time performance and maintenance efficiency of the digital clone knowledge base in the scenario of continuous document iteration.

[0194] In one embodiment, the system further performs abnormal scenario identification and rollback processing during each processing step, so as to perform rollback processing of the corresponding steps in abnormal scenarios. Specifically, abnormal scenario identification and rollback processing includes: If the parsing process corresponding to the parsing path selection result fails, the document will be switched to the next parsing path in descending order of the calculated score value to re-execute the document parsing; if all parsing paths fail, the document to be processed will be marked as a document to be reviewed. When the resolution confidence of an image in the document to be processed is lower than the preset quality threshold, structural resolution or hybrid resolution rollback is triggered; if the resolution confidence of the image is still lower than the preset quality threshold after rollback, the semantic evidence unit corresponding to the image is marked as a low confidence unit. When the candidate pre-filtering result is empty, directly generate an insufficient evidence knowledge response package and skip the subsequent response generation; When controlled loading fails to load any semantic evidence units within the remaining latency budget, prioritize outputting a degraded summary result or a rejection message. When the conflict penalties of multiple version states all exceed the preset limit, a version conflict warning is output, along with a publicly available summary of non-conflicting evidence. When permission verification fails, the corresponding field is masked, obscured, or generalized to a range, and insufficient permissions are written as the reason for downgrading into the knowledge response package. When the version of the digital avatar on the front end is inconsistent with the consistency identifier of the back end session, the front end should pause the subsequent response group broadcast and wait for the session reconstruction or version alignment to be completed.

[0195] In this embodiment, by performing abnormal scenario identification and corresponding rollback processing in the processing stages such as parsing, candidate pre-screening, controlled loading, version determination, permission verification and response output, the robustness, stability and feasibility of the digital clone system under various abnormal scenarios are significantly improved.

[0196] In one embodiment, the storage cluster in the system is used to store various types of data. The storage cluster employs, for example... Figure 7 The logical data structure shown stores relevant data records during the construction, updating, and retrieval of the evidence binding knowledge base. Each data record includes the following: Document registration records should include at least the document identifier, subject identifier, version identifier, parent version identifier, document format, file fingerprint, source channel, creation time, permission level, parsing path, and processing status. Location node records include at least one or more of the following: node identifier, document identifier, version identifier, page number or logical page number, block type, reading order, top-left corner coordinates, bottom-right corner coordinates, structure path, table path, cell path, comment path, image area path, character start offset, and character end offset. Semantic evidence unit records include at least the evidence unit identifier, document identifier, version identifier, start node identifier, end node identifier, evidence text, vector reference, time start point, time end point, time precision, permission tag, quality score, evidence hash, and review mark; Knowledge node records (i.e., node records in the fact graph) include at least the knowledge node identifier, node type, canonical name, alias set, node confidence, and normalized label; Knowledge edge records (i.e., relation edge records in the fact graph) should include at least the relation edge identifier, source node identifier, relation type, target node identifier, relation confidence, number of supporting evidence, and version status identifier. Version status records should include at least the status identifier, object identifier, effective start time, effective end time, document version identifier, support strength, source priority, and conflict flag. The knowledge response package record includes at least the response package identifier, query identifier, response text, evidence identifier set, address set, permission mask, version snapshot, style tag, response confidence, and degradation status.

[0197] The data structure of the storage cluster in the system can be implemented using relational databases, document databases, graph databases, or combinations thereof, but is not limited to any specific database product.

[0198] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention. It should be noted that the information interaction, execution process, etc. between the above devices / units are based on the same concept as the method embodiments of the present invention, and their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here.

[0199] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the functions described above can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0200] This invention also provides an electronic device, such as... Figure 8 As shown, the electronic device includes: at least one processor, a memory, a communication interface, and a computer program stored in the memory and executable on at least one of the processors. When the processor executes the computer program, it implements the steps in any of the above-described method embodiments, or when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments. For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0201] Those skilled in the art will understand that Figure 8 The electronic device described is merely an example and does not constitute a limitation on the electronic device. It may include more or fewer components than shown, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0202] The aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0203] The memory can be an internal storage unit of the electronic device, such as a hard drive or RAM. The memory can also be an external storage device of the electronic device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units of the electronic device.

[0204] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, enables an electronic device to implement the above-described method or the functions of the above-described system.

[0205] This invention provides a computer program product that, when executed on an electronic device, enables the electronic device to implement the above-described method or the functions of the above-described system.

[0206] If the integrated unit is implemented as 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, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can at least include: a recording medium capable of carrying and transmitting computer program code, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0207] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0208] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0209] In the embodiments provided by this invention, it should be understood that the disclosed apparatus / device and method can be implemented in other ways. For example, the apparatus / device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0210] The units described as separate components may or may not be physically separate. 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 network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0211] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for constructing and training a digital clone knowledge base that supports multi-format document parsing, characterized in that, include: Receive at least one document to be processed, and perform format recognition, version identification, and preprocessing on the document to be processed; Document features are extracted from the preprocessed document to be processed, and at least one target parsing path is determined from multiple candidate parsing paths based on the document features of the document to be processed; the multiple candidate parsing paths include at least two of the following: structural parsing path, layout parsing path, image parsing path, and hybrid parsing path; The document to be processed is parsed based on the target parsing path to generate multiple basic content units of the document to be processed, as well as layout structure information or logical structure information corresponding to the basic content units. A source location address mapping that can point back to the source document location is generated for the basic content units. The source location address mapping is used for one or more of the following: evidence reference, evidence continuity judgment, version comparison, or answer source writing back. Based on one or more of the following: content boundaries, structural boundaries, semantic boundaries, temporal boundaries, or event boundaries, multiple basic content units are segmented, merged, or aggregated, and corresponding source location address mappings are attached to construct address-based semantic evidence units. Knowledge extraction, cross-document alignment, version organization, and index construction are performed on the semantic evidence units to form an evidence binding knowledge base. The evidence binding knowledge base includes one or more of the following: a set of semantic evidence units, knowledge objects associated with the semantic evidence units, a version organization structure, and a retrieval index. Based on the evidence-binding knowledge base, the knowledge layer and style layer of the digital clone are trained separately, and the trained digital clone is run to perform tasks.

2. The method according to claim 1, characterized in that, The method further includes: Receive change documents, deletion instructions, or permission change information for updating the evidence binding knowledge base, wherein the change documents include newly added documents and updated documents; Based on the hierarchical fingerprint recognition change set involving the object, and performing differential reconstruction, update or invalidation processing on one or more of the semantic evidence units, knowledge objects, version status, permission tags or index items corresponding to the change set, a new version of the evidence binding knowledge base is formed to train the digital clone.

3. The method according to claim 2, characterized in that, The differential reconstruction includes: generating hierarchical fingerprints based on the changed document at the paragraph level, table level, cell level, chapter level, page level, event level, field level, or address path level; identifying the change set based on the generated hierarchical fingerprints; and reconstructing, updating, or invalidating one or more of the semantic evidence units, knowledge objects, version status, permission tags, and index items corresponding to the change set. Specifically, when an object in the change set has an associated object with which it has a continuity or dependency relationship, the object in the change set and the associated object are reconstructed together; the associated object includes adjacent content of the same document, as well as associated units of the same knowledge object, the same event chain, or the same version state.

4. The method according to claim 1, characterized in that, The method further includes: Perform one or more of the following on the digital clone: ​​parameter calibration, configuration, or strategy optimization: Specifically, based on one or more of the following factors during the operation of the digital clone: ​​address back pointer error, time version selection error rate, permission pruning error rate, response latency exceeding the limit rate, and the proportion of evidence not covered in the response, one or more parameters of parsing weight, boundary threshold, version arbitration weight, controlled loading weight, and response threshold are calibrated and updated.

5. The method according to claim 1, characterized in that, The method further includes: When a query or call request is received during the operation of the digital clone, query constraint information is generated based on the query or call request. The query constraint information includes one or more of the following: time constraint, permission constraint, evidence constraint, style constraint, context budget constraint, and latency budget constraint. Based on the query constraint information, candidate filtering, recall sorting, controlled loading, and permission pruning are performed on the semantic evidence units in the evidence binding knowledge base to obtain the loaded evidence set; A knowledge response package is generated based on the loaded evidence set, and one of the following is output according to the evidence coverage status, permission status, version status or confidence status of the knowledge response package: normal response, downgraded response, rejection prompt or pending review prompt. Specifically, when the response confidence level is lower than a preset threshold, the evidence coverage rate is lower than a preset threshold, the permission verification fails, or the version conflict cannot be arbitrated, a downgraded response, a rejection prompt, or a pending review prompt will be output.

6. The method according to claim 5, characterized in that, The process of performing candidate filtering, recall sorting, controlled loading, and permission pruning on the semantic evidence units in the evidence binding knowledge base yields a loaded evidence set, including: Based on the query time constraint, the permission constraint, and the evidence quality threshold, multiple semantic evidence units are pre-filtered to obtain a candidate set; The candidate semantic evidence units in the candidate set are mixed for recall and sorting to obtain the recalled evidence set; Loading constraints are applied to the recalled evidence set. Based on the loading constraint results and the recall score, candidate semantic evidence units are selected from the recalled evidence set. The selected candidate semantic evidence units are then jointly loaded based on the evidence continuity rule to obtain the loaded evidence set. The loading constraints include one or more of context budget, latency budget, permission constraints, and evidence quantity constraints; the recall score is determined by at least two of semantic similarity, keyword matching degree, graph correlation degree, time matching degree, source weight, permission satisfaction degree, evidence completeness, conflict penalty, and evidence continuity; the evidence continuity rules include one or more continuity relationships among the same source document, adjacent address paths, the same table, the same list, the same chapter, the same event chain, or the same version state.

7. The method according to claim 5, characterized in that, The knowledge response package includes answer content and response control information; the response control information includes one or more of the following: evidence address set, evidence identifier, permission mask, version snapshot, style tag, response confidence, downgrade status, session consistency identifier, group sequence number, and pending review mark; when the response control information includes the evidence address set or evidence identifier, the evidence address set or evidence identifier is associated with at least one semantic evidence unit in the loaded evidence set.

8. The method according to claim 7, characterized in that, The knowledge response package can be divided into multiple response groups, each response group including a response fragment and response control sub-information. The response control sub-information includes one or more of the following: evidence address subset, evidence identifier subset, permission mask subset, version snapshot, style tag, time sequence number, session consistency identifier, group confidence, and pending review mark.

9. The method according to claim 8, characterized in that, Before each round of dialogue begins or before the response group is output, a session consistency identifier is generated. The session consistency identifier is determined based on one or more of the following: front-end image version, knowledge version, style version, permission version, model version, and round number, and is written into at least one response group of the same round of dialogue.

10. The method according to claim 7, characterized in that, The knowledge response package or response group is output to the dialogue call terminal, knowledge question and answer call terminal or knowledge service interface of the digital avatar; the dialogue call terminal of the digital avatar includes a front-end call terminal constructed based on one or more of the following: reference image, reference video, three-dimensional image parameters, visual character parameters, voice parameters, action parameters or text interaction parameters. The dialogue invocation client is used to make real-time dialogue invocations based on one or more of the response group's time sequence number, session consistency identifier, permission mask, version snapshot, or style tag.

11. The method according to any one of claims 1 to 10, characterized in that, The step of generating a source location address mapping for the basic content unit that can point back to the source document location includes: Based on the basic content units and the page layout information or logical structure information, the reading order of the document to be processed is reconstructed to obtain the reading order number of at least some of the basic content units; Based on the registration record of the document to be processed and the reading order number, a source location address mapping that can point back to the source document location is generated for the basic content unit. The source location address mapping includes the physical address, logical address, or a combination of both of the basic content unit.

12. The method according to claim 11, characterized in that, The reading order reconstruction includes: determining the relative position, column consistency, style continuity, sequence number continuity, and overlap relationship between the basic content units based on the page structure information or logical structure information, and determining the reading order number of the basic content units based on one or more of the relative position, column consistency, style continuity, sequence number continuity, and overlap relationship between the basic content units.

13. The method according to any one of claims 1 to 12, characterized in that, The document features are used to characterize the parsability, structure, or layout complexity of the document to be processed; the multiple candidate parsing paths include one or more of the following: structure parsing path, layout parsing path, image parsing path, table parsing path, presentation parsing path, webpage parsing path, annotation parsing path, rule parsing path, model parsing path, or hybrid parsing path. The basic content units include at least one of the following: paragraph blocks, heading blocks, list blocks, table blocks, cell blocks, field blocks, annotation blocks, and text blocks corresponding to image areas; the document features include one or more of the following: raw text extraction rate, image conversion ratio, table density, multi-column probability, layout complexity, style hierarchy strength, encoding reliability, formula density, number of blocks within a page, logical structure strength, footnote / annotation density, number of attachments, document source information, and historical parsing results.

14. The method according to any one of claims 1 to 12, characterized in that, The source location address mapping includes the document identifier and location information for locating the original content location. The location information includes one or more of the following: version identifier, content unit identifier, page number or logical page number, normalized coordinates, reading order number, content unit type, character start and end offset, structure path, field path, table path, cell path, annotation path, attachment path, and image area path.

15. The method according to any one of claims 1 to 12, characterized in that, The semantic evidence unit is determined based on one or more of the following: boundary score, combined score, or rule-based determination of the basic content unit; wherein, the boundary score is determined based on one or more of the following: title level change, semantic similarity change, page layout change, time boundary change, and block type change; the combined score is determined based on one or more of the following: title or table header similarity, column structure similarity, serial number continuity, contextual continuity, and pagination relationship. Furthermore, a quality score and evidence hash are calculated for each semantic evidence unit, and semantic evidence units with quality scores lower than a preset quality threshold are marked as low-confidence units or units to be reviewed; when the confidence level of any of the parsing results in text parsing, table parsing, layout parsing, image parsing or mixed parsing is lower than the preset threshold, one or more of the following processes are triggered: backtracking parsing, supplementary parsing or manual review.

16. The method according to any one of claims 1 to 12, characterized in that, The cross-document alignment includes: calculating the alignment score of two candidate knowledge objects from different documents based on name similarity, contextual semantic similarity, temporal overlap, role consistency, and source consistency; and mapping the two candidate knowledge objects from different documents to the same knowledge object when the alignment score is greater than or equal to a preset score threshold.

17. The method according to any one of claims 1 to 12, characterized in that, The version organization structure includes a time version chain, which records the set of version states of the same knowledge object under different document versions, different effective times, and different supporting evidence. Based on query time constraints, evidence support strength, source authority, evidence consistency, and conflict penalty, multiple version states in the set of version states are arbitrated to select the target version state as the basis for the answer.

18. The method according to any one of claims 1 to 12, characterized in that, When simultaneously training the knowledge layer and the style layer, the knowledge layer and the style layer are trained separately by one or more of the following methods: orthogonal constraint, gating isolation, mutual information minimization, projection isolation, parameter isolation, cue isolation, sample isolation, or output verification isolation, so as to reduce the coupling between knowledge representation and style representation. The knowledge layer training is used to optimize one or more of the following: evidence matching, recall ranking, version consistency, reference location, evidence coverage, permission verification, and rejection judgment. The style layer training is used to optimize one or more of the following: terminology preference, sentence structure preference, length preference, narrative perspective, output structure, and grouping expression.

19. The method according to claim 18, characterized in that, During training, structured noise sampling is performed on multiple semantic evidence units to perform knowledge layer training based on the sampled evidence; the structured noise sampling includes at least one or more of evidence deletion, similar evidence replacement, evidence masking, and cross-version replacement.

20. A system for constructing and training a digital clone knowledge base that supports parsing multiple document formats, characterized in that, include: The document access and version registration module is used to receive at least one document to be processed and perform format recognition, version identification and preprocessing on the document to be processed. The parsing routing module is used to extract document features from the preprocessed document to be processed, and determine at least one target parsing path from multiple candidate parsing paths based on the document features of the document to be processed; the multiple candidate parsing paths include one or more of the following: structure parsing path, layout parsing path, image parsing path, table parsing path, presentation parsing path, webpage parsing path, annotation parsing path, rule parsing path, model parsing path, or hybrid parsing path; The structure parsing and address mapping module is used to parse the document to be processed based on the target parsing path, generate multiple basic content units of the document to be processed, as well as layout structure information or logical structure information corresponding to the basic content units, and generate source location address mappings for the basic content units that can point back to the source document location. The source location address mappings are used for one or more of the following: evidence reference, evidence continuity judgment, version comparison, or answer source writing back. The semantic evidence unit construction module is used to perform segmentation, merging or aggregation on multiple basic content units based on one or more of content boundaries, structural boundaries, semantic boundaries, time boundaries or event boundaries, and attach corresponding source location address mappings to construct address-based semantic evidence units. The knowledge extraction and alignment module is used to perform knowledge extraction and cross-document alignment on the semantic evidence unit to obtain the knowledge object associated with the semantic evidence unit. The knowledge base organization and version indexing module is used to perform version organization and index construction on the semantic evidence unit to form an evidence binding knowledge base. The evidence binding knowledge base includes one or more of the following: a set of semantic evidence units, knowledge objects associated with the semantic evidence units, a version organization structure, and a retrieval index. The hierarchical training module is used to separate and train the knowledge layer and style layer of the digital clone based on the evidence binding knowledge base, so as to run the trained digital clone to perform tasks. A storage cluster is used to store one or more of the following: the basic content unit, the source location address mapping, the semantic evidence unit, the knowledge object, the version organization structure, and the retrieval index.

21. The system according to claim 20, characterized in that, Also includes: The incremental update and evaluation module is used to identify the set of changes based on the hierarchical fingerprint of the objects involved when receiving a changed document, deletion instruction or permission change information, and to perform differential reconstruction, update or invalidation processing on one or more of the semantic evidence units, knowledge objects, version status, permission tags or index items corresponding to the set of changes, so as to form a new version of the evidence binding knowledge base to train the digital clone. The changed document includes newly added documents and updated documents.

22. The system according to claim 20, characterized in that, Also includes: The query orchestration module is used to generate query constraint information based on the query or call request when a query or call request is received during the operation of the digital clone. The query constraint information includes one or more of the following: time constraint, permission constraint, evidence constraint, style constraint, context budget constraint, and latency budget constraint. The candidate pre-filtering and controlled loading module is used to perform candidate filtering, recall sorting and controlled loading on the semantic evidence units in the evidence binding knowledge base according to the query constraint information, so as to obtain the loaded evidence set. A knowledge response package generation module is used to generate a knowledge response package based on the loaded evidence set; The permission and security control module is used to perform permission pruning on the semantic evidence unit, candidate fragment, or response fragment based on the query constraint information when generating the loaded evidence set, the knowledge response package, or the response group. The permission pruning supports at least one or more granularities of document level, chapter level, paragraph level, table level, field level, evidence level, or response group level. The digital clone dialogue interface module is used to output one of the following based on the evidence coverage status, permission status, version status, or confidence status of the knowledge response package: normal response, downgraded response, rejection prompt, or pending review prompt.

23. An electronic device comprising a memory, a processor, a communication interface, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it causes the electronic device to perform the method described in any one of claims 1 to 19, or to perform the function of the system described in any one of claims 20 to 22.

24. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it causes the electronic device to perform the method described in any one of claims 1 to 19, or to perform the function of the system described in any one of claims 20 to 22.

25. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, causes an electronic device to perform the method described in any one of claims 1 to 19, or to perform the function of the system described in any one of claims 20 to 22.

Images