A verifiable reference and conflict arbitration multi-agent context control system and method

Abstract

The application discloses a kind of verifiable reference and conflict arbitration multi-agent context control system and method, system includes: multi-source data access layer, analysis slice index layer, module context control layer, evidence management layer, generation and export layer, interaction and playback layer.Method steps include: establishing task record, analysis and cache material, context controller filters material set, on-demand call vector retrieval and networking search, Agent generates initial draft and supports re-search, assert evidence alignment and find minimum sufficient support, conflict arbitration and scoring, permission check and desensitization, generate final document and export, playback component recovers context snapshot.The scheme generates through multi-agent cooperation and blueprint-driven chapter-level parallel, combined with assertion evidence alignment, conflict automatic arbitration and permission whole-process propagation, realizes the verifiable, traceable and safe compliance of generated content;Built-in re-search and local rewriting closed loop triggered by proofreading, significantly improve the quality and generation efficiency of long report.

Description

Technical Field

[0001] This invention belongs to the fields of financial technology, artificial intelligence application, information retrieval and generative content production technology, and specifically relates to a multi-agent context control system and method with verifiable citations and conflict arbitration. Background Technology

[0002] PEVC (Private Equity and Venture Capital) due diligence requires in-depth analysis of massive amounts of heterogeneous materials (business plans, meeting minutes, public opinion, industry databases, etc.) and the generation of reports with verifiable evidence. While the Retrieval-Augmented Generation (RAG) framework addresses knowledge updates and source citation issues, it primarily focuses on general question-and-answer formats and lacks assertion-level fine-grained evidence binding specific to due diligence scenarios. Although studies like PaperTrail introduce claim-evidence grounding, they fail to address information conflicts and authority differences among multi-source heterogeneous materials (company self-reports, public information, interview transcripts). Multi-agent debate-based aggregation methods like MADAM-RAG still rely on general question-and-answer formats and lack credibility stratification, temporal priority, and authority inheritance mechanisms for investment decision-making. Publicly available patents such as US20240346256A1 and CN119671744A involve RAG response generation, Agent orchestration, or financial report generation, but none of them form a complete solution integrating context control, assertion-level citation, conflict arbitration, permission propagation, process evidence storage, and asynchronous export.

[0003] The existing technology has the following main problems: the granularity of citation is too coarse, and key assertions lack direct evidence support; silent side selection in the event of multi-source conflicts leads to the dilution of risk information; the boundaries of material use between modules are blurred, making it easy to misuse across modules; the generation process lacks evidence fingerprints and action chain records, which cannot meet the requirements for audit playback; permissions and de-identification tags are interrupted in the retrieval, generation, and export stages; and long-chain processes lack cache reuse and asynchronous orchestration, resulting in insufficient performance and stability.

[0004] Therefore, the above problems urgently need to be solved. Summary of the Invention

[0005] Purpose of the Invention: To overcome the above shortcomings, the purpose of this invention is to provide a multi-Agent context control system and method for verifiable citations and conflict arbitration. Through multi-Agent collaboration and blueprint-driven chapter-level parallel generation, combined with assertion evidence alignment, automatic conflict arbitration, and full-process permission propagation, the generated content is verifiable, traceable, and secure and compliant. The built-in review-triggered re-retrieval and partial rewriting closed loop significantly improves the quality and generation efficiency of long reports.

[0006] Technical Solution: To achieve the above objectives, this invention provides a multi-agent context control system capable of verifying references and arbitrating conflicts, the system comprising at least: The multi-source data access layer is used to receive project-uploaded files, meeting minutes, interview records, industry knowledge base results, and online search results. The parsing slice index layer is used for unified parsing, structured slicing, metadata extraction, and vectorized indexing of various types of data; The module context control layer is used to dynamically determine the set of documents, industry libraries, network tools, and agents that can be called based on the report type, project stage, chapter task, and user question; The evidence governance layer is used to perform evidence binding, source mapping, conflict detection, conflict arbitration, permission inheritance, and de-identification control on key assertions in the generated text; The generation and export layer is used to assemble the evidence-verified content into the target format and upload and store it, while saving the exported version and task status. The interaction and playback layer is used to push the generation progress, preview content, download link, and playback location information to the front end.

[0007] The entire process, from data access to final export, forms a closed loop, with the evidence governance layer built into the generation process. This distinguishes it from the traditional RAG system's shortcomings of "generating directly after retrieval without verification." Meanwhile, the module context control layer dynamically adjusts available resources based on report type, stage, and task to prevent irrelevant materials from contaminating the generated content, thereby improving the relevance of the generated content and token utilization. Furthermore, the interaction and playback layer pushes progress and playback location in real time, solving the pain points of no user feedback and inability to trace the generation process in long task generation.

[0008] Furthermore, the module context control layer also includes: The blueprint planning component is used to generate chapter blueprints based on report type. It describes chapters in a tree structure and forms a directed acyclic graph of tasks with dependency edges. The scheduler executes sub-chapter tasks concurrently based on dependencies and then triggers parent chapter tasks, which significantly shortens the generation time of long reports. The agent scheduling component includes at least a planning agent, a file extraction agent, a business analysis agent, an industry analysis agent, a public opinion agent, and a publishing agent. The responsibilities of multiple agents are separated, which facilitates individual debugging, replacement, or granting different model / tool ​​permissions, thereby improving system flexibility and scalability. The evidence governance layer also includes: The assertion evidence alignment component is used to extract key assertions from the generated content and bind evidence fragment information and credibility scores; unlike document-level citation, the assertion evidence alignment component binds key sentences to the smallest evidence fragment, improving citation accuracy and verifiability; The conflict arbitration component is used to sort conflicting evidence from multiple sources by source reliability, timeliness, entity consistency, cross-source consistency, and access control integrity, and output three types of results: "accepted," "disputed," and "awaiting manual review." When conflicting evidence from multiple sources is presented, it no longer simply discards or blindly accepts it, thus balancing automation efficiency with decision security. The permission label propagation component is used to attach access levels, sensitivity levels, and de-identification strategies to document fragments, and propagates them synchronously with the evidence chain during the recall, writing, citation, and export processes. Permissions are propagated along with the evidence chain to prevent permission leakage during the recall or citation process, thus meeting the compliance requirements of scenarios such as financial due diligence and compliance reporting.

[0009] Furthermore, the conflict arbitration component first implements hard gating based on the consistency of permissions and entities. If no usable evidence exists or there is strong rebuttal that cannot be resolved, the output is "Pending manual review." If the overall credibility of the best evidence reaches a threshold and there is no equally credible rebuttal, the output is "Accepted." If multiple highly credible pieces of evidence have similar scores or the conflict cannot be automatically resolved, the output is "Dispute Remains Unresolved," and a conflict list and review recommendations are generated. Hard gating prioritizes filtering out evidence without permissions or with mismatched entities before weighted calculation, ensuring security and reducing invalid calculations. It clarifies the decision boundary, using "leading advantage reaching a threshold" as the acceptance condition to avoid ambiguous rulings. When scores are close, the output is "Dispute Remains Unresolved," and a conflict list is generated to guide manual review rather than automatic acceptance of errors. At the same time, the output includes a conflict list and review recommendations, making the arbitration process transparent and meeting the needs of auditing and appeals.

[0010] Furthermore, when determining whether evidence can be displayed, the permission label propagation component checks at least the tenant project scope, user role access level, report type suitability, sensitivity level threshold, anonymization rule implementation, and source permission policy. If all conditions are met, the original text is allowed to be displayed; otherwise, partial anonymization, summary replacement, or only the conclusion is retained without exposing the original text is performed. It checks not only user roles but also report types and sensitivity level thresholds to prevent highly sensitive materials from being cited in low-level reports. A tiered downgrade strategy is designed: when the original text is prohibited, anonymization is attempted first, followed by summary replacement, and finally, the conclusion is retained without exposing the original text, maximizing the preservation of factual value while strictly adhering to compliance standards.

[0011] This invention also discloses a multi-agent context control method for implementing verifiable references and conflict arbitration in the system, comprising the following steps: S1): Receive project information, user-uploaded files, chapter outlines, or user questions, and create task records; S2): Perform unified analysis of materials, establish a cache based on content hashing, and extract metadata and searchable text; S3): The module context controller filters the set of materials that are allowed to enter the current context window based on module type, project stage, and permission scope; S4): If the current chapter relies on industry knowledge or publicly available information, then vector retrieval and network search will be invoked respectively to form a local evidence set and an external evidence set; S5): The Agent scheduling component generates a draft based on chapter tasks, allowing for on-demand retrieval and binding of chapter evidence sets; S6): The assertion evidence alignment component extracts key assertions from the initial draft and searches for the minimum sufficient supporting fragment in the candidate evidence set. If no sufficient support is found, it is marked as "low confidence" or "needs further evidence". S7): The conflict arbitration component scores and generates interpretations of conflicting evidence clusters; S8): The permission tag propagation component checks whether the accepted evidence is allowed to be displayed. If not, it triggers desensitization or summary replacement. S9): The publishing component merges the chapter content to generate the final document and outputs reference materials and evidence index; S10): The export component converts the document to the specified format and uploads it for storage, writing back the download address; S11): The replay component restores the context snapshot at the time of generation based on the task ID and the exported version number.

[0012] Furthermore, step S11, the replay component, can restore the context snapshot at the time of generation, facilitating review, appeal, or comparative iteration; step S3 filters materials before generation, step S6 binds assertions and evidence during generation, and step S7 arbitrates conflicts after generation, ensuring that the output content is verifiable and conflicts have been resolved; external knowledge is introduced as needed, step S4 determines whether to call vector retrieval or network search based on chapter dependencies, avoiding unnecessary API calls and costs; step S8 checks permissions again before exporting to prevent leakage caused by loss of permission tags in previous steps.

[0013] Furthermore, in step S3, a hard-gating screening is first performed, and then the candidate materials are comprehensively scored and ranked according to relevance, reliability, timeliness, coverage, diversity, and conflict penalty. Under the constraint of token budget, the output material set is truncated, and conflicting materials are marked as "reserved for dispute / pending manual review". Conflicting evidence is not directly discarded, but rather marked and entered into the arbitration process, thus retaining the negative information needed for decision-making.

[0014] Furthermore, in step S5, after the initial draft is generated, the review component outputs structured diagnostics. If any of the following triggering conditions are met—insufficient coverage, lack of evidence for assertions, unsupported citations, low search quality, non-arbitrable conflicts, or failure to meet timeliness requirements—a re-search request is automatically generated. After obtaining supplementary evidence, it is merged and partially rewritten, iterating until the quality threshold is met or the budget limit is reached. Automatic review after generation, triggering supplementary searches upon detection of deficiencies, eliminates the need for manual intervention to determine "whether further searching is needed," thus improving quality. Simultaneously, only relevant paragraphs are rewritten, retaining content that has already passed validation, saving tokens and avoiding the introduction of new errors.

[0015] Furthermore, the minimum sufficient support mentioned in step S6 refers to: first filtering evidence by entity and time window, and then judging the degree of support through semantic implication or numerical consistency; aiming at minimizing the number of tokens and fragments, prioritizing the selection of evidence where a single fragment meets the sufficient support threshold, otherwise combining 2-3 fragments until the sufficient support threshold is reached; if it still does not reach the threshold, it is marked as low confidence or requires supplementary evidence and a re-search is triggered. Defining "sufficient support" as reaching the sufficient support threshold avoids the subjective judgment of relevance by LLM; at the same time, an active supplementary evidence mechanism is designed, which automatically triggers a re-search after marking as low confidence or requiring supplementary evidence, forming an automated closed loop of "assertion verification → supplementary evidence → rewriting".

[0016] As can be seen from the above technical solution, the present invention has the following beneficial effects: 1. This invention provides a multi-agent context control system and method for verifiable citations and conflict arbitration. Through assertion evidence alignment and conflict arbitration mechanisms, each key assertion is bound to a minimum sufficient evidence fragment, and multi-source conflicts are automatically arbitrated (accepted / reserved for dispute / awaiting manual review). Combined with process evidence preservation and replay components, the retrieval parameters, model version, and agent action chain at the time of generation can be fully restored, enabling content tracing and audit review.

[0017] 2. The present invention provides a multi-Agent context control system and method for verifiable citations and conflict arbitration. The module context control layer dynamically filters the callable materials and agent set based on report type, project stage and permission scope. The permission tag is propagated throughout the evidence chain and automatically performs desensitization or summary replacement in the recall, writing and export stages to ensure compliance while avoiding irrelevant materials from contaminating the quality of the generated data.

[0018] 3. This invention provides a multi-agent context control system and method for verifiable references and conflict arbitration. The blueprint planning component decomposes the report into a directed acyclic graph of tasks, supporting parallel execution of sub-sections before triggering the parent section summary, significantly shortening the generation time of long reports. The review component automatically diagnoses issues such as insufficient coverage and non-arbitrable conflicts, triggering re-retrieval and partial rewriting, iterating to the quality threshold, forming a closed-loop self-healing mechanism. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the architecture of a multi-agent context control system with verifiable references and conflict arbitration as described in this invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention. Example

[0021] In this embodiment, as Figure 1 This invention discloses a multi-agent context control system capable of verifying references and arbitrating conflicts, the system comprising at least: The multi-source data access layer is used to receive project-uploaded files, meeting minutes, interview records, industry knowledge base results, and online search results. The parsing slice index layer is used for unified parsing, structured slicing, metadata extraction, and vectorized indexing of various types of data; The module context control layer is used to dynamically determine the set of documents, industry libraries, network tools, and agents that can be called based on the report type, project stage, chapter task, and user question; The evidence governance layer is used to perform evidence binding, source mapping, conflict detection, conflict arbitration, permission inheritance, and de-identification control on key assertions in the generated text; The generation and export layer is used to assemble the evidence-verified content into the target format and upload it for storage, while saving the exported version and task status. Specifically, a long task queue can be maintained through Redis / Celery to store Markdown to Word, Word to PDF, and PDF upload objects. The interaction and playback layer is used to push generation progress, preview content, download links, and playback location information to the front end; specifically, this can be achieved through SSE / WebSocket / task query interfaces.

[0022] In this embodiment, the module context control layer further includes: The blueprint planning component is used to generate chapter blueprints based on report type. Chapters are described in a tree structure and task directed acyclic graphs are formed with dependency edges. The scheduler executes sub-chapter tasks concurrently based on dependency relationships and then triggers parent chapter tasks. The agent scheduling component includes at least a planning agent, a file extraction agent, a business analysis agent, an industry analysis agent, a public opinion agent, and a publishing agent; The evidence governance layer also includes: The assertion evidence alignment component is used to extract key assertions from the generated content and bind evidence fragment information and credibility levels; The key assertions are generated directly by LLM based on prompts; the credibility level is directly given by LLM as high / medium / low; the evidence fragment information includes evidence fragment ID, source name, timestamp, page number or chapter number.

[0023] Specifically, for the large model portion, a pre-trained, instruction-fine-tuned / aligned large language model is preferred as the base model. For specific business needs, this system does not retrain the model parameters; instead, it injects external evidence into the context through retrieval-enhanced generation to achieve verifiable generation.

[0024] The conflict arbitration component is used to sort conflicting evidence from multiple sources by source reliability, timeliness, entity consistency, cross-source consistency, and access control integrity, and output three types of results: "accepted, disputed, and pending manual review". Specifically, the system combines hard gating, the highest score, the difference between the highest and second-highest scores, and the intensity of the conflict to map three types of results: When the highest-scoring candidate passes the gating of permissions, entities, and time windows, and its score reaches the acceptance threshold, is 10% higher than the second-highest candidate, and there is no equally credible rebuttal, the output is "accepted"; when the highest-scoring candidate is available but the score difference with other high-scoring candidates is insufficient or there is a high-credibility conflict that cannot be automatically resolved, the output is "dispute remains"; when the candidate fails the hard gating, the highest score is lower than the review threshold, or there is a strong rebuttal conflict that makes the system unable to arbitrate stably, the output is "pending manual review".

[0025] The permission tag propagation component is used to attach access levels, sensitivity levels, and de-identification strategies to document fragments, and propagate them synchronously with the evidence chain during recall, writing, citation, and export processes.

[0026] Specifically, in the conflict arbitration component, source reliability is derived from the source authority classification table combined with historical verification calibration; timeliness is calculated by the difference between the disclosure / update time of the evidence and the chapter time window through a decay function; entity consistency is obtained by entity identification and unified ID / alias matching (semantic similarity as a fallback); cross-source consistency indicates whether a candidate piece of evidence is consistent with other independent sources in this task; and permission integrity is jointly determined by the completeness of the permission tag when the evidence is added to the database and the current user access verification. If the permission is not met, the evidence is directly removed from the arbitration.

[0027] In this embodiment, the conflict arbitration component first performs hard gating based on the consistency of permissions and entities; if there is no available evidence or there is a strong rebuttal that cannot be resolved, it outputs "awaiting manual review"; if the overall credibility of the best evidence reaches the threshold and there is no equally credible rebuttal, it outputs "accepted"; if the scores of multiple highly credible pieces of evidence are close or the conflict cannot be automatically resolved, it outputs "dispute remains" and generates a conflict list and review suggestions.

[0028] In this embodiment, when determining whether evidence can be displayed, the permission label propagation component checks at least the tenant project scope, user role access level, report type compatibility, sensitivity level threshold, desensitization rule execution status, and source permission policy. If all conditions are met, the original text is allowed to be displayed; otherwise, partial desensitization, summary replacement, or only the conclusion is retained without exposing the original text is performed.

[0029] In this embodiment, the present invention also discloses a multi-Agent context control method for implementing verifiable references and conflict arbitration in the system, comprising the following steps: S1): Receive project information, user-uploaded files, chapter outlines, or user questions, and create task records; S2): Perform unified analysis of materials, establish a cache based on content hashing, and extract metadata and searchable text; Specifically, the content hash is generated based on the source type (source_type), file ID (file_id), file name (file_name), and source address (source_url) to avoid duplicate parsing; S3): The module context controller filters the set of materials that are allowed to enter the current context window based on module type, project stage, and permission scope; S4): If the current chapter relies on industry knowledge or publicly available information, then vector retrieval and network search will be invoked respectively to form a local evidence set and an external evidence set; Specifically, the logic for determining dependence in step S4 is as follows: A priori determination is made based on the evidence source dependence label in the chapter blueprint; if the chapter is marked as requiring industry knowledge bases or public information, vector retrieval and online search are invoked respectively to form two types of evidence sets. If there is no priori label, after local retrieval, gating indicators such as coverage, retrieval similarity, source diversity, timeliness, and conflict intensity are used to determine whether local evidence is insufficient or unarbitrable. If any triggering condition is met, the chapter is determined to depend on industry knowledge or public information, and the corresponding external retrieval process is initiated.

[0030] S5): The Agent scheduling component generates a draft based on chapter tasks, allowing for on-demand retrieval and binding of chapter evidence sets; S6): The assertion evidence alignment component extracts key assertions from the initial draft and searches for the minimum sufficient supporting fragment in the candidate evidence set. If no sufficient support is found, it is marked as "low confidence" or "needs further evidence". S7): The conflict arbitration component scores and generates interpretations of conflicting evidence clusters; S8): The permission tag propagation component checks whether the accepted evidence is allowed to be displayed. If not, it triggers desensitization or summary replacement. S9): The publishing component merges the chapter content to generate the final document and outputs reference materials and evidence index; S10): The export component converts the document to the specified format and uploads it for storage, writing back the download address; S11): The replay component restores the context snapshot at the time of generation based on the task ID and the exported version number.

[0031] Preferably, the scoring function used in S7 for scoring can be expressed as: Score = α·R + β·T + γ·E + δ·C + ε·P Where R represents source reliability, T represents timeliness, E represents entity consistency, C represents cross-source consistency, P represents access control integrity, and α, β, γ, δ, and ε are the weights of the above scores, respectively. Specifically, the weights are set using a "module default configuration + rule-based adaptive adjustment" approach. The system first presets a set of basic weights based on the module type, and then dynamically adjusts them according to the characteristics of the current conflict cluster. These characteristics include at least: conflict intensity, overall timeliness of evidence, degree of entity ambiguity, source dispersion, and historical verification performance. For example, when different sources give significantly contradictory results for the same indicator, the cross-source consistency weight is increased; when a chapter requires recent information but the candidate evidence is generally older, the timeliness weight is increased; when there is confusion due to entities with the same name or related companies, the entity matching weight is increased; when high-authority sources and low-authority sources coexist, the source reliability weight is increased. After adjustment, each weight is normalized, with permission integrity selection used as a hard gate; evidence that does not meet the permission conditions is directly eliminated and does not participate in subsequent scoring. In this embodiment, in step S3, a hard-gating screening is first performed, and then the candidate materials are ranked by comprehensive scoring based on relevance, reliability, timeliness, coverage, diversity and conflict penalty. Under the token budget constraint, the output material set is truncated in the Top-K and hierarchical summary manner, and conflicting materials are marked as "reserved for dispute / awaiting manual review".

[0032] Specifically, the coverage object is the set of information requirements predefined in the current chapter, such as market size, growth rate, competitive landscape, etc.; the degree of support of candidate materials for each requirement item is recorded as the coverage value, and the material coverage or material set coverage is calculated accordingly.

[0033] Specifically, the conflict penalty is used to quantify the degree of inconsistency between candidate materials and selected high-confidence materials on the same entity, the same time window, the same indicator, or the same event. It can be weighted and summed according to the conflict type and conflict intensity, and is used as a deduction item in the comprehensive score.

[0034] Specifically, diversity represents the gain of candidate materials on the current set of selected materials in terms of source type, source instance, temporal distribution, and semantic redundancy, which can be calculated by "new source bonus + low redundancy bonus".

[0035] Specifically, the comprehensive scoring and ranking optimization adopts a method of hard gating followed by greedy iterative selection. That is, the initial score is calculated first based on relevance, reliability, timeliness, coverage, diversity and conflict penalty. Then, under the token budget constraint, materials are selected step by step according to the marginal contribution score. When the budget is exceeded, a hierarchical summary is used to replace the original text output.

[0036] Specifically, the token budget constraint refers to setting a maximum token limit for candidate materials entering the context window when generating a single chapter. This limit is obtained by subtracting system prompts, chapter task descriptions, historical states, output reserves, and safety margins from the total context capacity of the model. For example: Assuming a certain chapter is titled "Market Size and Competitive Landscape," the total context capacity of the model is set as follows: Bmodel=32000 The following are reserved: System prompt: Bsys=2000 Chapter Task Description: Binst=1500 Historical status: Bhist=2500 Output reservation: Bout=6000 Safety margin: Bsafe=1000 The current chapter's material budget is: Bctx=32000−2000−1500−2500−6000−1000=19000 That is, this chapter allows a maximum of approximately 19,000 tokens of material to enter the context window.

[0037] Specifically, the hard gating in step S3 includes permission scope, entity consistency, allowed material types for modules / chaps, time windows, deduplication, and version control.

[0038] In this embodiment, in step S5, after the initial draft is generated, the review component outputs a structured diagnosis. If any of the following triggering conditions occur: insufficient coverage, no evidence assertion, unsupported citation, low retrieval quality, non-arbitrable conflict, or failure to meet timeliness, a re-retrieval request is automatically generated. After obtaining supplementary evidence, the results are merged and partially rewritten. The process iterates until the quality threshold is met or the budget limit is reached.

[0039] In this embodiment, the minimum sufficient support mentioned in step S6 refers to: first filtering evidence by entity and time window, and then judging the degree of support by semantic implication or numerical consistency; with the goal of minimizing the number of tokens and fragments, prioritizing the selection of evidence that a single fragment meets the sufficient support threshold, otherwise combining 2 to 3 fragments until the sufficient support threshold is reached; if it still cannot be reached, it is marked as low confidence or pending supplementary evidence and triggers a re-search.

[0040] Specifically, the time window refers to the effective time range corresponding to the current assertion or section, such as a fiscal year, the most recent period, the past year, or up to a certain date. The system extracts the reporting period, disclosure date, publication time, or update time of candidate evidence and matches it with this time range. Only evidence whose time falls within this range or whose deviation does not exceed the tolerance threshold is allowed to enter the support judgment, while other evidence is eliminated or downweighted.

[0041] Specifically, in step S6, whether sufficient support is achieved is determined by gating combined with a threshold: first, evidence with no authorization / inconsistent entities / mismatched time windows is filtered out; then, an assertion-evidence support score is calculated for the available evidence. The support score is calculated by weighting semantic implication, numerical consistency, and relevance, and preferably combines source reliability, timeliness, entity matching degree, and cross-source consistency. If a single fragment or a small number of fragment combinations cause the support score to reach the threshold and there is no equally credible rebuttal evidence, it is judged as sufficient support. If it only reaches the low credibility threshold, it is marked as low credibility. If it is below the low credibility threshold or there is no available evidence, it is marked as needing supplementary evidence and triggers a re-search.

[0042] Specifically, the credible rebuttal evidence refers to evidence fragments that simultaneously pass gating such as permissions, entity consistency, and time window matching, and can form a clear semantic rebuttal or numerical conflict to the target assertion, and whose source reliability, timeliness, entity matching degree, and cross-source consistency reach a preset credible threshold; if the credible rebuttal strength of the rebuttal evidence is close to or reaches the support strength of the best supporting evidence, it is considered as "equally credible rebuttal evidence".

[0043] Specifically, it fully supports thresholds greater than the low confidence threshold.

[0044] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A multi-agent context control system capable of verifying references and arbitrating conflicts, characterized in that, The system includes at least: The multi-source data access layer is used to receive project-uploaded files, meeting minutes, interview records, industry knowledge base results, and online search results. The parsing slice index layer is used for unified parsing, structured slicing, metadata extraction, and vectorized indexing of various types of data; The module context control layer is used to dynamically determine the set of documents, industry libraries, network tools, and agents that can be called based on the report type, project stage, chapter task, and user question; The evidence governance layer is used to perform evidence binding, source mapping, conflict detection, conflict arbitration, permission inheritance, and de-identification control on key assertions in the generated text; The generation and export layer is used to assemble the evidence-verified content into the target format and upload and store it, while saving the exported version and task status. The interaction and playback layer is used to push the generation progress, preview content, download link, and playback location information to the front end.

2. The system according to claim 1, characterized in that, The module context control layer also includes: The blueprint planning component is used to generate chapter blueprints based on report type. Chapters are described in a tree structure and task directed acyclic graphs are formed with dependency edges. The scheduler executes sub-chapter tasks concurrently based on dependency relationships and then triggers parent chapter tasks. The agent scheduling component includes at least a planning agent, a file extraction agent, a business analysis agent, an industry analysis agent, a public opinion agent, and a publishing agent; The evidence governance layer also includes: The assertion evidence alignment component is used to extract key assertions from the generated content and bind evidence fragment information and credibility scores; The conflict arbitration component is used to sort conflicting evidence from multiple sources by source reliability, timeliness, entity consistency, cross-source consistency, and access control integrity, and output three types of results: "accepted, disputed, and pending manual review". The permission tag propagation component is used to attach access levels, sensitivity levels, and de-identification strategies to document fragments, and propagate them synchronously with the evidence chain during recall, writing, citation, and export processes.

3. The system according to claim 2, characterized in that, The conflict arbitration component first performs hard gating based on the consistency of permissions and entities; if there is no available evidence or there is a strong rebuttal that cannot be resolved, it outputs "Pending manual review"; if the overall credibility of the best evidence reaches the threshold and there is no equally credible rebuttal, it outputs "Accepted"; if the scores of multiple highly credible evidences are close or the conflict cannot be automatically resolved, it outputs "Dispute Remains Unresolved" and generates a conflict list and review suggestions.

4. The system according to claim 2, characterized in that, When determining whether evidence should be displayed, the permission label propagation component checks at least the tenant project scope, user role access level, report type compatibility, sensitivity level threshold, desensitization rule execution, and source permission policy. If all conditions are met, the original text is allowed to be displayed; otherwise, partial desensitization, summary replacement, or only the conclusion is retained without exposing the original text is performed.

5. A multi-agent context control method for implementing the system of any one of claims 1 to 4 with verifiable references and conflict arbitration, characterized in that, Includes the following steps: S1): Receive project information, user-uploaded files, chapter outlines, or user questions, and create task records; S2): Perform unified analysis of materials, establish a cache based on content hashing, and extract metadata and searchable text; S3): The module context controller filters the set of materials that are allowed to enter the current context window based on module type, project stage, and permission scope; S4): If the current chapter relies on industry knowledge or publicly available information, then vector retrieval and network search will be invoked respectively to form a local evidence set and an external evidence set; S5): The Agent scheduling component generates a draft based on chapter tasks, allowing for on-demand retrieval and binding of chapter evidence sets; S6): The assertion evidence alignment component extracts key assertions from the initial draft and searches for the minimum sufficient supporting fragment in the candidate evidence set. If no sufficient support is found, it is marked as "low confidence" or "needs further evidence". S7): The conflict arbitration component scores and generates interpretations of conflicting evidence clusters; S8): The permission tag propagation component checks whether the accepted evidence is allowed to be displayed. If not, it triggers desensitization or summary replacement. S9): The publishing component merges the chapter content to generate the final document and outputs reference materials and evidence index; S10): The export component converts the document to the specified format and uploads it for storage, writing back the download address; S11): The replay component restores the context snapshot at the time of generation based on the task ID and the exported version number.

6. The method according to claim 5, characterized in that, In step S3, a hard-gating screening is first performed, and then the candidate materials are ranked by comprehensive scoring based on relevance, reliability, timeliness, coverage, diversity and conflict penalty. The output material set is truncated under the token budget constraint, and conflicting materials are marked as "reserved for dispute / awaiting manual review".

7. The method according to claim 5, characterized in that, In step S5, after the initial draft is generated, the review component outputs a structured diagnosis. If any of the following triggering conditions are met, such as insufficient coverage, lack of evidence for assertion, unsupported citations, low search quality, non-arbitrable conflict, or failure to meet timeliness, a re-search request is automatically generated. After obtaining supplementary evidence, the results are merged and partially rewritten. The process iterates until the quality threshold is met or the budget limit is reached.

8. The method according to claim 5, characterized in that, The minimum sufficient support mentioned in step S6 refers to: first filtering evidence by entity and time window, and then judging the degree of support by semantic implication or numerical consistency; with the goal of minimizing the number of tokens and fragments, prioritizing the selection of evidence that meets the sufficient support threshold for a single fragment, otherwise combining 2 to 3 fragments until the sufficient support threshold is reached; if it still cannot be reached, it is marked as low confidence or pending supplementary evidence and a re-search is triggered.

Images