Proposal review method and system based on evidence verification and multi-agent cooperation

Abstract

The invention relates to the technical field of artificial intelligence multi-agents, in particular to a proposal review method and system based on evidence verification and multi-agent cooperation. A cross validation mechanism based on a review information pool is introduced. The cross validation mechanism is a multi-agent round table cooperation mechanism, and a preliminary review result generated by an agent is not directly output, but passes multiple rounds of mutual verification, supplementation and correction until a preset convergence condition is met. According to the evidence verification process, cognitive deviation and knowledge blind areas possibly existing in a single agent are effectively eliminated, and the omission ratio and the misjudgment rate of the review result are remarkably reduced, so that the accuracy, the integrity and the reliability of a final review report are ensured; the technical problems of low efficiency and low accuracy caused by insufficient review depth, single angle and lack of verification in the prior art are effectively solved, and the review efficiency and decision quality of enterprise project proposals are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence multi-agent technology, and in particular to a proposal review method and system based on evidence verification and multi-agent collaboration. Background Technology

[0002] In the process of enterprises promoting project proposals such as intelligent products or algorithm models, the review process involves multi-dimensional evaluation, especially in areas such as the use of personal data, ethical review of generated content, and market positioning. This typically requires round-by-round evaluations from multiple departments, including legal, compliance, and marketing. Current practices largely rely on manual processes, which are not only inefficient but also struggle to cope with increasingly complex compliance requirements, severely impacting the response speed and decision-making efficiency of proposals.

[0003] Current AI-powered review systems primarily focus on automating general approval workflows and configuring rules for document review processes, lacking the ability to deeply understand the semantics of proposals. These systems struggle to accurately interpret complex legal clauses and industry terminology within contracts or proposals, leading to missed checks or misjudgments during compliance reviews. Furthermore, most current systems employ a single, large-scale AI model for review, failing to meet the multi-faceted review needs of proposal evaluation, including legal interpretation and rule verification, resulting in insufficient accuracy in intelligent review. Therefore, improving the deep understanding of proposals by intelligent review systems and achieving accurate multi-dimensional proposal review has become a pressing technical challenge. Summary of the Invention

[0004] This invention aims to provide a proposal review method and system based on evidence verification and multi-agent collaboration, so as to improve the agent's ability to deeply understand the content of the proposal, realize multi-angle compliance review of project proposals, and enhance the comprehensiveness and completeness of the artificial intelligence proposal review results.

[0005] To achieve the above objectives, the first aspect of the present invention provides a proposal review method based on evidence verification and multi-agent collaboration, comprising the following steps: Obtain the proposal file and perform field validation and data cleaning on the proposal file to obtain the proposal dataset; Semantic parsing and semantic extraction are performed on the proposed dataset to obtain key point vectors and keyword sets; Based on the key point vector and the keyword set, several target intelligent agents are matched in the preset intelligent agent library, and then each target intelligent agent is linked to the corresponding review database. For any of the target intelligent agents: The target intelligent agent is input with the key point vector and the keyword set, so that the target intelligent agent can retrieve a number of candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence; Based on the verification score of each candidate review evidence, a number of basic review evidence that meet the preset verification score threshold are selected from the number of candidate review evidence. The target intelligent agent generates basic review information based on the key point vector, the keyword set, and several basic review evidences, and uses the several basic review evidences and the basic review information as a set of basic review texts for the target intelligent agent; Generate several types of basic review texts based on several types of target intelligent agents; A review information pool is constructed based on several types of basic review texts; Several target intelligent agents are cross-validated several times according to the review information pool until any cross-validation in any round satisfies the preset convergence condition. The review information pool is then updated based on the result of the cross-validation in that round. A structured review report is generated based on the review information pool.

[0006] The aforementioned proposal review method based on evidence verification and multi-agent collaboration first performs semantic parsing and semantic extraction on the proposal documents, transforming the unstructured proposal text content into structured semantic representations such as key point vectors and keyword sets. This enables subsequent agents to accurately capture the core intent and complex connotations of the proposal, thereby greatly improving the agents' ability to deeply understand the semantic content of the proposal.

[0007] Secondly, this invention constructs a multi-agent collaborative review framework that can dynamically match and invoke specialized agents from different fields such as law, compliance, and market based on the semantic features of the proposal, and links each agent to its own dedicated review database. This design overcomes the limitations of a single general-purpose large-scale model agent in terms of professional knowledge, ensuring that the review process can be carried out in parallel from multiple professional dimensions, achieving comprehensive coverage of review perspectives and professionalism in review opinions.

[0008] Furthermore, this invention introduces a weighted scoring and threshold screening mechanism for evidence. By conducting multi-dimensional quantitative evaluation of candidate review evidence and rigorously screening it according to preset verification score thresholds, a verification technology logic of verification before review is constructed. This design ensures that all basic review evidence entering the subsequent target agent review process has high accuracy and reliability, avoiding the use of outdated, inaccurate, or irrelevant database information that could lead to inaccurate generated review texts. In addition, this mechanism strictly constrains the agent's review process within the framework of already verified review database evidence, effectively avoiding unfounded inferences and information illusions, laying a solid factual foundation for subsequent cross-validation among multiple agents, and significantly improving the accuracy and reliability of generated review opinions.

[0009] Finally, this invention introduces a cross-validation mechanism based on a review information pool. This cross-validation mechanism is a multi-agent round-table collaboration mechanism, where the initial review results generated by the agents are not directly output, but are verified, supplemented, and corrected through multiple rounds until a preset convergence condition is met. This effectively eliminates the cognitive biases and knowledge blind spots that may exist in a single agent, significantly reducing the missed detection rate and false judgment rate of the review results. This ensures the accuracy, completeness, and reliability of the final review report, effectively solving the technical problems of low efficiency and low accuracy caused by insufficient review depth, single perspective, and lack of verification in the prior art, and significantly improving the review efficiency and decision-making quality of enterprise project proposals.

[0010] Preferably, the step of weighting and scoring each candidate piece of review evidence to obtain a verification score for each candidate piece of review evidence includes: The target agent is input with the key point vector and the keyword set so that the target agent can retrieve several candidate review evidences in the corresponding review database, and then analyze the content relevance, precedent similarity, version recentity and source credibility of each candidate review evidence. Based on the content relevance, precedent similarity, version recentity, and source credibility of each candidate review evidence, the verification score of each candidate review evidence is analyzed. This implementation method employs multi-dimensional quantitative analysis and weighted scoring of candidate review evidence based on its content relevance, precedent similarity, version recentity, and source credibility. This allows for comprehensive consideration of the evidence's intrinsic relevance, practical guidance value, timeliness, and reliability, thereby selecting fundamental review evidence that highly aligns with the proposal content, possesses strong reference value, and is authentic and reliable. This not only significantly improves the accuracy and objectivity of the evidence verification process but also provides solid data support for the subsequent generation of high-quality fundamental review information by the intelligent agent, ensuring the rigor and accuracy of the generated review text from the outset.

[0011] It should be noted that the content relevance is derived by calculating the semantic similarity between the key vector of the proposal and the vectorized representation of the candidate review evidence, in order to quantify the degree of fit between the evidence text and the proposal topic, ensuring that the evidence cited by the agent is highly related to the content of the proposal at the semantic level; the precedent similarity is calculated by comparing the structured feature vectors of the proposal and the candidate review evidence, which may include technical fields, business models, project scale, etc., aiming to screen out similar precedents with practical reference value from the perspective of project profile; the version recentity is a quantitative indicator derived by parsing the timestamp in the metadata of the candidate review evidence, assigning a higher indicator value to review evidence with a later publication date, ensuring the timeliness of the review basis; the source credibility is the result of quantitatively evaluating the publication source of the candidate review evidence based on a preset source authority scoring table or a trained credibility classification model. This scoring table or model is based on features such as the domain name type and historical accuracy of the source to filter low-quality or false information, ensuring that the review evidence input into the agent has high reliability.

[0012] Further, the step of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The intelligent agent library includes several pre-trained candidate intelligent agents, including a legal interpretation intelligent agent, a rule verification intelligent agent, a case retrieval intelligent agent, and a contingency plan retrieval intelligent agent; Among them, the review database corresponding to the legal interpretation intelligent agent is a legal clause database, which stores legal clause data; The review database corresponding to the rule verification intelligent agent is a compliance list template database, which stores industry rule text data. The review database corresponding to the case retrieval agent is a historical case database, which stores judicial judgment case document data. The review database corresponding to the plan retrieval agent is a plan template database, which stores executable proposal file data.

[0013] In this implementation, a library of specialized candidate agents, including those for legal interpretation, rule verification, case retrieval, and preliminary plan retrieval, is pre-set. Target agents are dynamically matched based on the semantic features of the proposals. This ensures that each review task is performed by an "expert" agent with relevant domain knowledge. This not only guarantees that each agent can conduct in-depth mining and analysis within its own area of ​​expertise, avoiding interpretation biases caused by knowledge generalization in general models, but also allows the generated review text to be linked to objective information in the corresponding review database as verification evidence, thus improving the traceability of the review text's viewpoints.

[0014] Further, the step of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The agent library includes several pre-trained candidate agents. For any candidate agent: Obtain the review responsibility text data and review domain feature tags of the candidate agent, and then construct agent metadata based on the review responsibility text data and the review domain feature tags; Calculate the text matching score and domain matching score between the key point vector, the keyword set, and the agent metadata, and then obtain the comprehensive matching score corresponding to the candidate agent based on the text matching score and the domain matching score; Obtain the comprehensive matching score corresponding to each candidate agent, and then select the candidate agents whose comprehensive matching scores meet the preset score conditions as the target agents.

[0015] In this implementation, metadata containing review responsibility text data and review domain feature labels is constructed for each candidate agent. The text matching score and domain matching score between the proposal's key point vector, keyword set, and this metadata are comprehensively calculated to form a quantified comprehensive matching score, which serves as the screening criterion for selecting suitable target agents for the target proposal document. This method not only identifies the literal correlation between the proposal text and the agent's responsibilities but also semantically identifies the domain-specific correlation between the proposal document and the agent's corresponding database. This ensures that each selected target agent is the most relevant and professional "expert" for the current proposal content, avoiding information noise and resource waste caused by irrelevant agents participating in the review. It achieves a dynamic agent combination driven by proposal semantics, providing a high-quality execution subject for subsequent multi-agent collaborative review, ultimately enhancing the accuracy, comprehensiveness, and completeness of the AI ​​proposal review results.

[0016] Further, the step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of cross-validation, for any target agent: Based on the review information pool, the target agent is input with the basic review text corresponding to any other target agent, so that the target agent can retrieve several candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence. Based on the verification score of each candidate review evidence, several second review evidences that meet the preset verification score threshold are selected from the several candidate review evidences. The target intelligent agent generates second review information based on the basic review text and several pieces of second review evidence, and uses the several pieces of second review evidence and the second review information as a set of second review texts corresponding to the basic review text; The basic review texts corresponding to the other target intelligent agents are analyzed respectively to obtain several second review texts corresponding to the basic review texts; The review information pool is updated based on several of the second review texts.

[0017] This implementation employs a multi-agent, round-table collaborative cross-validation mechanism. Each agent performs a secondary review of the opinions expressed in the review texts output by the other agents, enabling mutual scrutiny and iterative correction among the agents. This achieves self-optimization and deep integration of the multi-agent review results. In this mechanism, each target agent does not complete the review in isolation but analyzes the basic review texts generated by all other agents and forms new second review texts based on these analyses. This method constructs a dynamic and mutually inspiring round-table collaborative argumentation environment. By analyzing and commenting on the review texts generated by other agents, it incentivizes the agent to respond to the basic review texts, generating new opinions and introducing new database evidence. This improves the quality of the review information in the review information pool, facilitating the acquisition of more comprehensive and complete review reports.

[0018] Further, the step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of the cross-validation described above: Semantic similarity clustering is performed on several second review texts to obtain several review text sets; For any set of review texts, calculate the mutual exclusion degree of review positions and the mutual exclusion degree of cited data for the set of review texts, and then obtain the degree of disagreement of review texts based on the mutual exclusion degree of review positions and the mutual exclusion degree of cited data, thereby obtaining the degree of disagreement of review texts corresponding to each set of review texts. If the divergence of all the review texts meets the preset divergence threshold, then the cross-validation round is determined to meet the convergence condition.

[0019] In this implementation, the second review texts are first clustered based on semantic similarity to facilitate subsequent analysis of the degree of disagreement among review texts belonging to the same topic. Then, the invention calculates mutual exclusion from two dimensions: review stance and cited data. Review stance mutual exclusion is used to analyze the disagreement of viewpoints or conclusions within the review texts themselves, while cited data mutual exclusion reflects the mutual exclusion between the database evidence information linked to by the review texts. The combined degree of disagreement in the review texts constitutes a comprehensive and profound quantitative indicator that objectively reflects the degree of disagreement within each viewpoint camp. Finally, convergence is only determined when the degree of disagreement of all review text sets is below a preset threshold, ensuring that the cross-validation loop does not prematurely terminate when there are significant unresolved disputes among agents, thereby guaranteeing the internal consistency of various viewpoints in the final review information pool and the integrity of the cited database evidence chain.

[0020] A second aspect of the present invention provides a proposal review system based on evidence verification and multi-agent collaboration, comprising: The proposal preprocessing module is used to perform the following steps: obtain the proposal file, and perform field validation and data cleaning on the proposal file to obtain the proposal dataset; perform semantic parsing and semantic extraction on the proposal dataset to obtain the key point vector and keyword set; The agent matching module is used to match several target agents in a preset agent library based on the key point vector and the keyword set, and then link each target agent to the corresponding review database. The agent initial evaluation module is used to perform the following steps: For any of the target intelligent agents: The target intelligent agent is input with the key point vector and the keyword set, so that the target intelligent agent can retrieve a number of candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence; Based on the verification score of each candidate review evidence, a number of basic review evidence that meet the preset verification score threshold are selected from the number of candidate review evidence. The target intelligent agent generates basic review information based on the key point vector, the keyword set, and several basic review evidences, and uses the several basic review evidences and the basic review information as a set of basic review texts for the target intelligent agent; Generate several types of basic review texts based on several types of target intelligent agents; The multi-agent cross-validation module is used to perform the following steps: A review information pool is constructed based on several types of basic review texts; several target agents are cross-validated several times according to the review information pool until any round of cross-validation meets a preset convergence condition, and the review information pool is updated based on the result of that round of cross-validation. The review report generation module is used to generate structured review reports based on the review information pool.

[0021] The aforementioned proposal review method based on evidence verification and multi-agent collaboration firstly transforms unstructured proposal text into structured semantic representations such as key point vectors and keyword sets through semantic parsing and extraction of proposal documents. This enables subsequent agents to accurately capture the core intent and complex connotations of the proposal, thereby significantly improving their deep semantic understanding of the proposal content. Secondly, this invention constructs a multi-agent collaborative review framework that dynamically matches and invokes specialized agents from different fields such as law, compliance, and market based on the semantic features of the proposal, linking each agent to its dedicated review database. This design overcomes the limitations of a single general-purpose large-scale model agent in terms of professional knowledge, ensuring that the review process can proceed in parallel from multiple professional dimensions, achieving comprehensive coverage of review perspectives and professional review opinions. Furthermore, this invention introduces a cross-validation mechanism based on a review information pool. This cross-validation mechanism is a multi-agent round-table collaboration mechanism, ensuring that the initial review results generated by the agents are not directly output, but rather verified, supplemented, and corrected through multiple rounds until the preset convergence conditions are met. This evidence verification process effectively eliminates the cognitive biases and knowledge blind spots that may exist in a single intelligent agent, significantly reducing the rate of missed detections and misjudgments in the review results, thereby ensuring the accuracy, completeness and reliability of the final review report. It effectively solves the technical problems of low efficiency and low accuracy caused by insufficient review depth, single perspective and lack of verification in the existing technology, and significantly improves the review efficiency and decision quality of enterprise project proposals.

[0022] Further, the step of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The intelligent agent library includes several pre-trained candidate intelligent agents, including a legal interpretation intelligent agent, a rule verification intelligent agent, a case retrieval intelligent agent, and a contingency plan retrieval intelligent agent; Among them, the review database corresponding to the legal interpretation intelligent agent is a legal clause database, which stores legal clause data; The review database corresponding to the rule verification intelligent agent is a compliance list template database, which stores industry rule text data. The review database corresponding to the case retrieval agent is a historical case database, which stores judicial judgment case document data. The review database corresponding to the plan retrieval agent is a plan template database, which stores executable proposal file data.

[0023] In this implementation, a library of specialized candidate agents, including those for legal interpretation, rule verification, case retrieval, and preliminary plan retrieval, is pre-set. Target agents are dynamically matched based on the semantic features of the proposals. This ensures that each review task is performed by an "expert" agent with relevant domain knowledge. This not only guarantees that each agent can conduct in-depth mining and analysis within its own area of ​​expertise, avoiding interpretation biases caused by knowledge generalization in general models, but also allows the generated review text to be linked to objective information in the corresponding review database as verification evidence, thus improving the traceability of the review text's viewpoints.

[0024] Further, the step of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The agent library includes several pre-trained candidate agents. For any candidate agent: Obtain the review responsibility text data and review domain feature tags of the candidate agent, and then construct agent metadata based on the review responsibility text data and the review domain feature tags; Calculate the text matching score and domain matching score between the key point vector, the keyword set, and the agent metadata, and then obtain the comprehensive matching score corresponding to the candidate agent based on the text matching score and the domain matching score; Obtain the comprehensive matching score corresponding to each candidate agent, and then select the candidate agents whose comprehensive matching scores meet the preset score conditions as the target agents.

[0025] In this implementation, metadata containing review responsibility text data and review domain feature labels is constructed for each candidate agent. The text matching score and domain matching score between the proposal's key point vector, keyword set, and this metadata are comprehensively calculated to form a quantified comprehensive matching score, which serves as the screening criterion for selecting suitable target agents for the target proposal document. This method not only identifies the literal correlation between the proposal text and the agent's responsibilities but also semantically identifies the domain-specific correlation between the proposal document and the agent's corresponding database. This ensures that each selected target agent is the most relevant and professional "expert" for the current proposal content, avoiding information noise and resource waste caused by irrelevant agents participating in the review. It achieves a dynamic agent combination driven by proposal semantics, providing a high-quality execution subject for subsequent multi-agent collaborative review, ultimately enhancing the accuracy, comprehensiveness, and completeness of the AI ​​proposal review results.

[0026] Further, the step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of cross-validation, for any target agent: Based on the review information pool, the target agent is input with the basic review text corresponding to any other target agent, so that the target agent can retrieve several candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence. Based on the verification score of each candidate review evidence, several second review evidences that meet the preset verification score threshold are selected from the several candidate review evidences. The target intelligent agent generates second review information based on the basic review text and several pieces of second review evidence, and uses the several pieces of second review evidence and the second review information as a set of second review texts corresponding to the basic review text; The basic review texts corresponding to the other target intelligent agents are analyzed respectively to obtain several second review texts corresponding to the basic review texts; The review information pool is updated based on several of the second review texts.

[0027] This implementation employs a multi-agent, round-table collaborative cross-validation mechanism. Each agent performs a secondary review of the opinions expressed in the review texts output by the other agents, enabling mutual scrutiny and iterative correction among the agents. This achieves self-optimization and deep integration of the multi-agent review results. In this mechanism, each target agent does not complete the review in isolation but analyzes the basic review texts generated by all other agents and forms new second review texts based on these analyses. This method constructs a dynamic and mutually inspiring round-table collaborative argumentation environment. By analyzing and commenting on the review texts generated by other agents, it incentivizes the agent to respond to the basic review texts, generating new opinions and introducing new database evidence. This improves the quality of the review information in the review information pool, facilitating the acquisition of more comprehensive and complete review reports.

[0028] Further, the step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of the cross-validation described above: Semantic similarity clustering is performed on several second review texts to obtain several review text sets; For any set of review texts, calculate the mutual exclusion degree of review positions and the mutual exclusion degree of cited data for the set of review texts, and then obtain the degree of disagreement of review texts based on the mutual exclusion degree of review positions and the mutual exclusion degree of cited data, thereby obtaining the degree of disagreement of review texts corresponding to each set of review texts. If the divergence of all the review texts meets the preset divergence threshold, then the cross-validation round is determined to meet the convergence condition.

[0029] In this implementation, the second review texts are first clustered based on semantic similarity to facilitate subsequent analysis of the degree of disagreement among review texts belonging to the same topic. Then, the invention calculates mutual exclusion from two dimensions: review stance and cited data. Review stance mutual exclusion is used to analyze the disagreement of viewpoints or conclusions within the review texts themselves, while cited data mutual exclusion reflects the mutual exclusion between the database evidence information linked to by the review texts. The combined degree of disagreement in the review texts constitutes a comprehensive and profound quantitative indicator that objectively reflects the degree of disagreement within each viewpoint camp. Finally, convergence is only determined when the degree of disagreement of all review text sets is below a preset threshold, ensuring that the cross-validation loop does not prematurely terminate when there are significant unresolved disputes among agents, thereby guaranteeing the internal consistency of various viewpoints in the final review information pool and the integrity of the cited database evidence chain. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating a proposal review method based on evidence verification and multi-agent collaboration provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a proposal review system based on evidence verification and multi-agent collaboration. Detailed Implementation

[0031] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the following detailed descriptions are exemplary and intended to provide further detailed explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0032] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0033] Before describing this application in detail with reference to the accompanying drawings and embodiments, the terms and application scenarios involved in this application will first be explained.

[0034] This invention aims to provide a proposal review method and system based on evidence verification and multi-agent collaboration, so as to improve the agent's ability to deeply understand the content of the proposal, realize multi-angle compliance review of project proposals, and enhance the comprehensiveness and completeness of the artificial intelligence proposal review results.

[0035] Please refer to Figure 1 To achieve the above objectives, the first embodiment of the present invention provides a proposal review method based on evidence verification and multi-agent collaboration, comprising the following steps: S101. Obtain the proposal file and perform field validation and data cleaning on the proposal file to obtain the proposal dataset; S102. Perform semantic parsing and semantic extraction on the proposal dataset to obtain the key point vector and keyword set; S103. Based on the key point vector and the keyword set, match several target intelligent agents in the preset intelligent agent library, and then link each target intelligent agent to the corresponding review database. S104. For any of the target intelligent agents: The target intelligent agent is input with the key point vector and the keyword set, so that the target intelligent agent can retrieve a number of candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence; Based on the verification score of each candidate review evidence, a number of basic review evidence that meet the preset verification score threshold are selected from the number of candidate review evidence. The target intelligent agent generates basic review information based on the key point vector, the keyword set, and several basic review evidences, and uses the several basic review evidences and the basic review information as a set of basic review texts for the target intelligent agent; Generate several types of basic review texts based on several types of target intelligent agents; S105. Construct a review information pool based on several types of basic review texts; S106. Several target intelligent agents are cross-validated in several rounds according to the review information pool until any round of cross-validation meets the preset convergence condition. The review information pool is then updated based on the result of the cross-validation in that round. S107. Generate a structured review report based on the review information pool.

[0036] The aforementioned proposal review method based on evidence verification and multi-agent collaboration firstly transforms unstructured proposal text into structured semantic representations such as key point vectors and keyword sets through semantic parsing and extraction of proposal documents. This enables subsequent agents to accurately capture the core intent and complex connotations of the proposal, thereby significantly improving their deep semantic understanding of the proposal content. Secondly, this invention constructs a multi-agent collaborative review framework that dynamically matches and invokes specialized agents from different fields such as law, compliance, and market based on the semantic features of the proposal, linking each agent to its dedicated review database. This design overcomes the limitations of a single general-purpose large-scale model agent in terms of professional knowledge, ensuring that the review process can proceed in parallel from multiple professional dimensions, achieving comprehensive coverage of review perspectives and professional review opinions. Furthermore, this invention introduces a cross-validation mechanism based on a review information pool. This cross-validation mechanism is a multi-agent round-table collaboration mechanism, ensuring that the initial review results generated by the agents are not directly output, but rather verified, supplemented, and corrected through multiple rounds until the preset convergence conditions are met. This evidence verification process effectively eliminates the cognitive biases and knowledge blind spots that may exist in a single intelligent agent, significantly reducing the rate of missed detections and misjudgments in the review results, thereby ensuring the accuracy, completeness and reliability of the final review report. It effectively solves the technical problems of low efficiency and low accuracy caused by insufficient review depth, single perspective and lack of verification in the existing technology, and significantly improves the review efficiency and decision quality of enterprise project proposals.

[0037] In one possible embodiment, in a certain set of basic review texts, the basic review information does not carry corresponding basic review evidence. This basic review text is not weighted when the review information pool is updated; that is, it is removed and not used for subsequent updates to the review information pool, thereby improving the accuracy and reliability of the structured review report.

[0038] In one specific embodiment, step S101 is specifically a proposal receiving and standardization process. This process first obtains the original proposal document, which typically includes, but is not limited to, multi-dimensional information such as product objectives, data types, algorithms or models, target audiences or channels, launch regions, and compliance actions already taken. Subsequently, the system performs field validation and data cleaning operations on the proposal document. Field validation aims to check the completeness and validity of each key field according to preset rules, such as verifying whether the data types match and whether the launch region is a valid value. Data cleaning is responsible for standardizing the text, including unifying terminology, correcting formatting errors, and removing redundant information. After the above processing, the system generates a structured proposal dataset.

[0039] Preferably, the proposal dataset is organized in JSON (JavaScript Object Notation) format to ensure its machine readability and ease of subsequent processing. For key fields missing in the original proposal, the system does not leave them blank but uniformly marks them as "unknown," thus ensuring the consistency and integrity of the data structure and providing high-quality, standardized data input for subsequent semantic parsing steps.

[0040] In one specific embodiment, step S102 involves deep semantic parsing and key information extraction of the proposal dataset. During this process, the system invokes a pre-trained large language model, such as the QWEN3 model, to perform deep semantic analysis on the standardized proposal dataset. This model aims to accurately identify and extract core semantic elements from the proposals, specifically including: the core purpose of the proposal, the data types involved, compliance elements for personal information protection, generative artificial intelligence-related behaviors, and high-frequency risk words that may violate laws and regulations. For example, whether it involves sensitive data such as biometric information or facial data; whether it includes user consent mechanisms, notification obligations, and filing status; and whether it involves content generation, digital watermarking, or the use of special identifiers.

[0041] Finally, the system integrates the extracted key information into a structured keyword set K, and further transforms these elements and their interrelationships into a computer-processable point vector F. Through this step, the unstructured proposal text is transformed into a data representation with both semantic depth and structured features, laying a solid data foundation for subsequent accurate matching and professional review by intelligent agents.

[0042] Further, step S103, which involves matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The intelligent agent library includes several pre-trained candidate intelligent agents, including a legal interpretation intelligent agent, a rule verification intelligent agent, a case retrieval intelligent agent, and a contingency plan retrieval intelligent agent; Among them, the review database corresponding to the legal interpretation intelligent agent is a legal clause database, which stores legal clause data; The review database corresponding to the rule verification intelligent agent is a compliance list template database, which stores industry rule text data. The review database corresponding to the case retrieval agent is a historical case database, which stores judicial judgment case document data. The review database corresponding to the plan retrieval agent is a plan template database, which stores executable proposal file data.

[0043] In this implementation, a library of specialized candidate agents, including those for legal interpretation, rule verification, case retrieval, and preliminary plan retrieval, is pre-set. Target agents are dynamically matched based on the semantic features of the proposals. This ensures that each review task is performed by an "expert" agent with relevant domain knowledge. This not only guarantees that each agent can conduct in-depth mining and analysis within its own area of ​​expertise, avoiding interpretation biases caused by knowledge generalization in general models, but also allows the generated review text to be linked to objective information in the corresponding review database as verification evidence, thus improving the traceability of the review text's viewpoints.

[0044] In one specific embodiment, step S103 includes: An independent session instance is generated for each matched target agent, and the session instance is injected with a set of role instructions and tool calling permissions corresponding to the target agent type. Specifically, for the legal interpretation agent, permission to call the legal clause database is granted; for the rule verification agent, permission to call the compliance list template database is granted; for the case retrieval agent, permission to call the historical case database is granted; and for the contingency plan retrieval agent, permission to call the contingency plan template database is granted. According to the preset interleaved parallel strategy, the session instances of each target agent are activated sequentially at preset time intervals (e.g., 200 milliseconds) to effectively avoid the API call frequency limit triggered by too many concurrent requests while ensuring the real-time performance of the review task.

[0045] Further, step S103, which involves matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The agent library includes several pre-trained candidate agents. For any candidate agent: Obtain the review responsibility text data and review domain feature tags of the candidate agent, and then construct agent metadata based on the review responsibility text data and the review domain feature tags; Calculate the text matching score and domain matching score between the key point vector, the keyword set, and the agent metadata, and then obtain the comprehensive matching score corresponding to the candidate agent based on the text matching score and the domain matching score; Obtain the comprehensive matching score corresponding to each candidate agent, and then select the candidate agents whose comprehensive matching scores meet the preset score conditions as the target agents.

[0046] In this implementation, metadata containing review responsibility text data and review domain feature labels is constructed for each candidate agent. The text matching score and domain matching score between the proposal's key point vector, keyword set, and this metadata are comprehensively calculated to form a quantified comprehensive matching score, which serves as the screening criterion for selecting suitable target agents for the target proposal document. This method not only identifies the literal correlation between the proposal text and the agent's responsibilities but also semantically identifies the domain-specific correlation between the proposal document and the agent's corresponding database. This ensures that each selected target agent is the most relevant and professional "expert" for the current proposal content, avoiding information noise and resource waste caused by irrelevant agents participating in the review. It achieves a dynamic agent combination driven by proposal semantics, providing a high-quality execution subject for subsequent multi-agent collaborative review, ultimately enhancing the accuracy, comprehensiveness, and completeness of the AI ​​proposal review results.

[0047] In one specific embodiment, step S103 includes: Obtain the key point vector F and keyword set K of the proposal; wherein, the key point vector F includes a semantic vector F_sem and a structured label F_slot. The semantic vector F_sem is used to characterize the semantic features of the proposal in dimensions such as training data type, algorithm method, market positioning and legal domain. The structured label F_slot is used to characterize the key attributes in the proposal, such as whether it involves face recognition (is_face=true), whether it involves minors (is_minor=false), legal domain (region=CN), and whether it is generative artificial intelligence (is_genAI=true).

[0048] The keyword set K is a set of keywords or phrases extracted from the proposal, used to preserve literal signals, such as "facial recognition", "cross-border transmission", "advertising", "minors", etc. Each candidate agent r in the agent library is bound to two types of metadata: the first type is review responsibility text data desc(r), which describes the responsibilities and terminology of the candidate agent r through text, facilitating semantic matching with the key point vector F and the keyword set K; the second type is review domain feature labels triggerers(r) and clauses(r), which represent the triggering conditions of the candidate agent and the relevant legal clause clusters in the corresponding review database, and are used for policy matching with the key point vector F and the keyword set K.

[0049] For each candidate agent r in the agent library, a matching score (score(r)) is calculated based on the key vector F and the keyword set K; the formula for calculating the matching score (score(r)) is as follows: score(r) = α · text_match(F_sem, K, r) + β · policy_match(F_slot,K, r) + γ · risk_prior(r | F_slot) + δ · history_freq(r | F_sem); Where α, β, γ, and δ are preset weight coefficients; text_match(F_sem, K, r) is used to characterize the semantic and literal matching degree between candidate agent r and the proposal text and keyword set, which comprehensively considers "semantic similarity" and "keyword coverage"; policy_match(F_slot, K, r) is used to characterize the trigger compliance degree of the proposal with candidate agent r in the policy and regulatory dimension, which evaluates the coverage degree of the structured label F_slot and keyword set K of the trigger conditions and the matching degree of the clauses and legal domains based on the trigger conditions triggerers(r) and legal clauses clusters clauses(r) bound to candidate agent r, and weights high-priority regulatory words; risk_prior(r | F_slot) is the risk prior probability of candidate agent r determined based on the structured label F_slot; history_freq(r | F_sem) is the historical participation frequency of candidate agent r determined based on the semantic vector F_sem; Based on the calculated matching score(r) of each candidate agent, the top k candidate agents with the highest matching scores are selected as target agents, forming a target agent set R_sel; and a priority π(r) is assigned to each target agent based on the matching score(r). At the same time, the target intelligent agent set R_sel is verified and supplemented according to the preset participation constraints. For example, when a proposal is detected to involve biometrics, the "legal interpretation intelligent agent" and the "rule verification intelligent agent" are forcibly included in the target intelligent agent set R_sel. Link each selected target agent to its corresponding review database so that the target agent can access it when performing review tasks.

[0050] In a preferred embodiment, the method for constructing the review database includes: First, raw data is collected from multiple heterogeneous data sources, including but not limited to officially published indexes of legal provisions, publicly available databases of judicial precedents, compliance lists of industry standards, and internal contingency plan template libraries.

[0051] Then, based on the type and professional field of the data content, the preprocessed data is imported and constructed into several specialized review databases, specifically including: a legal clause database storing legal provisions for the legal interpretation agent; a compliance list template database storing industry rule provisions for the rule verification agent; a historical case database storing judicial judgment case documents for the case retrieval agent; and a plan template database storing executable proposal documents for the plan retrieval agent. Finally, for each specific proposal review task, the system uses the aforementioned key point vector and keyword set as search instructions to perform multi-channel parallel searches in the above-mentioned specialized review databases, forming a broad-spectrum candidate evidence library.

[0052] Finally, an evidence verification module is activated. This module quantifies and scores each candidate piece of evidence in the candidate evidence database based on a preset weighting function. The weighting factors of the weighting function include clause relevance, precedent similarity, version recentity, and source credibility. By setting a verification score threshold, the system filters out candidate evidence with scores higher than the threshold, combining them into a high-precision, highly relevant review database for the corresponding target agent to use in this review. This allows the target agent to initiate parallel arguments by performing legal interpretation, rule auditing, precedent retrieval, and risk planning according to an interleaved scheduling protocol, provided that the verified evidence constraints are met.

[0053] Furthermore, based on the review database with the aforementioned evidence verification constraints, in a specific embodiment, step S104 includes: For the aforementioned case retrieval agent, after receiving the key point vector and keyword set, it first dynamically generates one or more search query terms based on the proposal content, and then calls the retrieval interface of the historical case database linked to it to obtain relevant judicial judgment cases, regulatory notices and industry public opinion data; for each piece of external information retrieved, its metadata is extracted, including but not limited to source URL, title, publication time, content summary and capture timestamp, and the metadata is bound with the review opinions generated based on the information to form a basic review text for case retrieval with a chain of evidence.

[0054] For the legal interpretation agent, after receiving the key point vector and keyword set, it performs clause matching and mapping in the legal clause database, establishes an association between the core viewpoints in the proposal and specific legal provisions, and generates a mapping table containing legal clause identifiers, specific clause citations, and relevance scores to the proposal viewpoints. Finally, it outputs a basic review text for legal interpretation containing explicit legal citations, and a mapping table of {clause_id, quote, relevance∈[0,1]}, and cites each item in the report.

[0055] For the rule verification agent, after receiving the key point vector and keyword set, it calls the compliance list template database, compares the key points of the proposal with the preset industry rule items one by one, identifies potential compliance risk points, and generates basic review text containing compliance inspection conclusions and corresponding rule basis.

[0056] The intelligent agent for retrieving the contingency plan receives the key point vector and keyword set, calls the contingency plan template database, retrieves executable proposal templates that match the proposal scenario, and generates preliminary response strategies or optimization suggestions based on the templates as basic review text.

[0057] Further, step S106, which involves having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of cross-validation, for any target agent: The basic review texts corresponding to the other target intelligent agents are analyzed respectively to obtain several second review texts corresponding to the basic review texts; The review information pool is updated based on several of the second review texts.

[0058] This implementation employs a multi-agent, round-table collaborative cross-validation mechanism. Each agent performs a secondary review of the opinions expressed in the review texts output by the other agents, enabling mutual scrutiny and iterative correction among the agents. This achieves self-optimization and deep integration of the multi-agent review results. In this mechanism, each target agent does not complete the review in isolation but analyzes the basic review texts generated by all other agents and forms new second review texts based on these analyses. This method constructs a dynamic and mutually inspiring round-table collaborative argumentation environment. By analyzing and commenting on the review texts generated by other agents, it incentivizes the agent to respond to the basic review texts, generating new opinions and introducing new database evidence. This improves the quality of the review information in the review information pool, facilitating the acquisition of more comprehensive and complete review reports.

[0059] Further, step S106, which involves having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of the cross-validation described above: Semantic similarity clustering is performed on several second review texts to obtain several review text sets; For any set of review texts, calculate the mutual exclusion degree of review positions and the mutual exclusion degree of cited data for the set of review texts, and then obtain the degree of disagreement of review texts based on the mutual exclusion degree of review positions and the mutual exclusion degree of cited data, thereby obtaining the degree of disagreement of review texts corresponding to each set of review texts. If the divergence of all the review texts meets the preset divergence threshold, then the cross-validation round is determined to meet the convergence condition.

[0060] In this implementation, the second review texts are first clustered based on semantic similarity to facilitate subsequent analysis of the degree of disagreement among review texts belonging to the same topic. Then, the invention calculates mutual exclusion from two dimensions: review stance and cited data. Review stance mutual exclusion is used to analyze the disagreement of viewpoints or conclusions within the review texts themselves, while cited data mutual exclusion reflects the mutual exclusion between the database evidence information linked to by the review texts. The combined degree of disagreement in the review texts constitutes a comprehensive and profound quantitative indicator that objectively reflects the degree of disagreement within each viewpoint camp. Finally, convergence is only determined when the degree of disagreement of all review text sets is below a preset threshold, ensuring that the cross-validation loop does not prematurely terminate when there are significant unresolved disputes among agents, thereby guaranteeing the internal consistency of various viewpoints in the final review information pool and the integrity of the cited database evidence chain.

[0061] In one possible embodiment, the preset convergence condition described in step S106 includes satisfying at least one of the following: Reaching the maximum number of rounds: When the current round of cross-validation reaches the preset maximum value T, the convergence condition is considered met.

[0062] Evidence increment stagnation: When the number of new valid pieces of evidence added to the review information pool from the previous round of cross-validation to the current round is less than a preset threshold ε (for example, the default value is 1 piece), the convergence condition is determined to be met.

[0063] Disagreement degree below threshold: When the degree of disagreement Δ of all review text sets in the current round is lower than the preset degree of disagreement threshold τ (for example, the default value is 0.2), it is determined that the convergence condition is met; wherein, the degree of disagreement Δ is calculated based on the number of mutually exclusive viewpoints and the degree of evidence conflict.

[0064] Budget limit reached: When the execution time or computational cost of the entire review process reaches the preset budget limit, the convergence condition is met and cross-validation is forcibly terminated.

[0065] In one specific embodiment, step S106 includes: Initiate round-based cross-validation: Set the number of rounds of cross-validation to t=1..T, and use the basic review text generated by each target agent in the first round as the initial review information pool.

[0066] Perform targeted cross-validation: In each round of cross-validation, a pre-defined host agent performs the targeted cross-validation process; the host agent encapsulates the atomic claims produced by any expert agent in the previous round into a structured data packet and distributes it to the other expert agents.

[0067] Generate rebuttal and supplementary evidence information: The expert agent that receives the structured data packet evaluates the atomized claim based on its own knowledge base and the linked review database, and returns a review position containing "support", "oppose" or "conditionally support" according to a unified schema, as well as feedback information with newly added cited evidence or legal provisions; among them, feedback information without cited evidence is not weighted in subsequent scoring.

[0068] Perform divergence quantification and convergence decision-making: After collecting feedback from all expert agents regarding the same atomized claim, perform the following operations: Identifying Disagreements: The claims in the feedback information are atomicated or normalized, and claims with semantic similarity ≥ θ_sem = 0.80 are aggregated into the same normalized claim; under the same normalized claim, if there are positions of support and opposition, or significant differences in threshold, then a disagreement is judged to exist.

[0069] Quantifying Conflicts: For each differing normative claim, calculate its degree of disagreement Δ, whereby the formula for calculating the degree of disagreement Δ is: Δ= ws×stance_gap + we×evidence_conflict + wc×clause_mismatch; Among them, ws, we, and wc are the preset weights for position difference, evidence conflict, and clause inconsistency, respectively; stance_gap is the degree of difference in review positions, for example, the difference between supportive and opposing review positions can be counted as 0.5; evidence_conflict is the degree of conflict of cited evidence databases, for example, the conflict between legal or normative sources and industry blog sources can be counted as 0.4; clause_mismatch is the degree of difference in applicable legal clauses, for example, inconsistency in jurisdiction, effective date, or clause level can be counted as 0.1.

[0070] Preferably, the weights for positional differences, conflict of evidence, and inconsistency of clauses are preset according to the target business scenario of the proposal document to adapt to the proposal needs under different business scenarios. For business scenarios that have undergone multiple proposal reviews, machine learning can be performed based on historical review data to learn the optimal weights for positional differences, conflict of evidence, and inconsistency of clauses, thereby improving the accuracy of the degree of disagreement Δ.

[0071] Convergence criteria are determined by comparing the calculated divergence Δ with a preset divergence threshold. If the divergence Δ of all standard claims is lower than the threshold, the cross-validation round is deemed to have met the convergence criteria. If the divergence Δ of any standard claim is higher than or equal to the threshold, the cross-validation round is deemed not to have met the convergence criteria, and the next round of cross-validation or re-search is triggered.

[0072] Update the review information pool: For rounds that meet the convergence criteria, convergence is achieved based on a comprehensive score of "evidence score × clause fit × role weight." Specific strategies include: Aggregation: If the descriptions of the two parties overlap but can coexist under certain conditions, a conditional conclusion is generated.

[0073] Concession: When the difference between the comprehensive scores of one side and the other side is significant, the conclusion with the higher score is adopted, and the conclusion with the lower score is downgraded to a supplementary or restrictive condition.

[0074] Pending Decision: If the degree of disagreement Δ in the reviewed text is still high or the evidence from both sides is insufficient, it is marked as pending, and a targeted re-search or manual review is automatically triggered; Finally, the convergence conclusion, pending marks and new evidence generated in this round are updated to the review information pool.

[0075] In a preferred embodiment, after step S106, the method further includes: Based on the updated review information pool, a risk level assessment is performed on the proposal, specifically including: calculating the overall risk level (risk) of the proposal, using the following formula: risk = max_r (w_r · severity_r); Where r represents different risk dimensions, w_r is the weight of each risk dimension, and its value is determined by the credibility of each target agent's role during the review process; severity_r is the severity of each risk dimension, and the severity value of each risk dimension is evaluated by considering the following factors: ; Among them, the severity of the clause S clauses The result is derived from the mapping of legal provisions in step S104. The assessment is based on the red line level (high, medium, or low) of the matched legal provisions and their compatibility with the region and effective period of the proposal.

[0076] Data sensitivity S datas The assessment is based on whether the data processed in the proposal contains sensitive personal information, such as biometrics, health information, and location information, and the proportion thereof.

[0077] Population Vulnerability S audiences The assessment is based on whether the proposed service targets include minors or other special groups as defined by law.

[0078] Processing range S scopes The assessment is based on the amount of data involved in the proposal, its specific purpose (such as model training, online inference, marketing promotion, etc.), and whether there is any external public exposure.

[0079] Cross-border and external transfers (Scrosss): The assessment is based on whether the proposal involves cross-border data transfers or sharing with third parties.

[0080] Enforcement precedent S precedents The results are derived from the dynamic retrieval and evidence binding in step S104. The assessment is based on the relevance and time proximity of the retrieved relevant penalty cases, regulatory notices, or judicial precedents to the current proposal.

[0081] Control gap S gaps Based on the rules in step S104, verify the output of the intelligent agent, compare it with the compliance checklist template, and assess the proportion of missing items in key control measures such as informed consent, privacy policy disclosure, data protection impact assessment (DPIA), algorithm filing, and watermarking.

[0082] Currently, it effectively alleviates S mitigs The severity of each of the above items will be reduced accordingly based on the technical and administrative mitigation measures that have been implemented or are planned to be implemented in the proposal.

[0083] Based on the above calculations, the overall risk level of the proposal was finally determined and divided into three levels: "high", "medium" and "low".

[0084] Disagreement handling and report triggering: If there is a conflict between target agents during the review process and the conflict is not resolved, the disagreement panel will be triggered to display in the generated structured review report. The disagreement panel is used to compare and display the evidence chains of the conflicting parties and provide an entry point for secondary retrieval for manual review or further automated verification.

[0085] In a preferred embodiment, after step S106, the method further includes: Generate executable mitigation paths: Based on the updated review information pool, generate executable mitigation paths for risk items rated as "high" or "medium" risk levels to realize the transformation from risk assessment to specific actions.

[0086] The generation of the mitigation path specifically includes: Output a list of executable actions: For each medium- to high-risk item, generate an action list containing specific execution items. The list includes, but is not limited to: specific action items, suggested completion timeline, designated person in charge, and required reference materials or templates, such as algorithm filing application number, user consent form template, data watermarking strategy document, and target audience restriction instructions, etc.

[0087] Set review trigger conditions: Set clear review trigger conditions for each mitigation path to automatically prompt or trigger a new round of review process when specific changes occur in the proposal; the review trigger conditions include, but are not limited to: changes in the scope of data processing, service launch to a new region, or updates to relevant laws and regulations.

[0088] Integrating Risks and Mitigation Solutions: Finally, the identified risk items and the generated mitigation paths are structurally integrated to form a core report product that includes the correspondence between "risks and mitigation solutions". This core report is a key component of the structured review report described in step S107, ensuring that the review conclusions are operable and have continued effectiveness.

[0089] In one specific embodiment, step S107 includes: Invoking the summary agent: A pre-defined summary agent initiates the report generation process based on the updated review information pool.

[0090] Generate a three-part report based on the template: The summarizing agent uses a preset report generation template to fill the structured data in the review information pool into the corresponding chapters of the report, generating a structured review report containing the following three parts.

[0091] Conclusion Overview: This section summarizes and presents the core conclusions of this review, including the overall risk level of the proposal (high / medium / low), key compliance assessment results, and a summary of the main points of disagreement identified during cross-validation.

[0092] Risk grading and evidence: The risk level is listed in detail for each item, and a complete chain of evidence is bound to each risk item. The chain of evidence includes, but is not limited to: source URL, title, publication time, content summary and crawl timestamp; at the same time, this part also includes a mapping table of legal clauses directly related to the viewpoint, with the structure {clause_id, quote, relevance}; for viewpoints with disagreements, the degree of disagreement Δ value and the corresponding status flag are displayed.

[0093] Mitigation Path and Timeline: For risk items rated as "high" or "medium" risk level, output the corresponding executable mitigation path. The path includes an action list, which details the specific action items, the suggested completion timeline, the designated person in charge, the required reference materials (such as algorithm filing number, consent form template, etc.), and the preset review trigger conditions.

[0094] Output both structured and readable dual-version reports: The structured review report includes two formats: one is a machine-readable structured data format, such as JSON or XML, which facilitates system integration and secondary development; the other is a formatted natural language report, which is easy for users to read and review directly.

[0095] In a preferred embodiment, the above-mentioned proposal review method based on evidence verification and multi-agent collaboration, in step S103, can further optimize agent selection by combining learnable Prompt optimization with task graphs, specifically including: The mapping process of matching several target intelligent agents in a preset intelligent agent library based on the key point vector and the keyword set is modeled as a graph-constrained set covering problem.

[0096] First, using a large language model (LLM), relevant task nodes and risk nodes are located on a pre-built task graph based on the semantic content of the proposal.

[0097] Secondly, the graph constraint solver is invoked to solve the set coverage problem under the premise of satisfying the preset hard constraints such as legal domain, regulations and risk control, so as to minimize the weighted objective function of "number of roles and invocation cost" and thus obtain a minimum and sufficient set of target intelligent agents.

[0098] Finally, the entire selection process is driven by a learnable Prompt optimizer. This optimizer uses retrieved similar cases as few-shot examples and combines contextual information such as the region of the proposal and applicable regulations to dynamically generate the most user-friendly instruction format for large language models. This ensures accurate recall of key roles and improves the interpretability of the selection process. This optimization process includes the following two improvements: First, graph awareness: By bringing the structured nodes and regulatory evidence from the task graph to the generated Prompt, rather than relying on descriptive writing in pure natural language, the LLM's ability to understand complex constraints is improved.

[0099] Second, a learnable Prompt optimizer: driven by online or offline data, it dynamically selects the optimal Prompt template and calibrates its output confidence, thereby continuously improving the accuracy of key role recall and the stability of system operation.

[0100] In a preferred embodiment, the multi-round cross-validation roundtable mechanism described in step S106 can also be used to support multiple human-machine collaboration modes to adapt to review scenarios with different complexity and timeliness requirements, specifically including: Fully automatic mode: In this mode, the roundtable mechanism operates completely autonomously until the preset convergence conditions are met, and automatically generates the final review report. It is suitable for standardized, low-risk routine proposal reviews.

[0101] Multi-round semi-automatic mode: This mode introduces human intervention nodes during the roundtable mechanism. For example, after the first round of cross-validation, the system can pause and display a disagreement panel for human experts to mark evidence gaps or verify key legal clauses. The system then performs targeted supplementary searches or adjusts weights based on human input, and continues with subsequent rounds of validation, forming a closed loop of "machine initial judgment - human correction - machine re-validation," suitable for reviewing proposals with high complexity or uncertainty.

[0102] Single-round human collaboration mode: In this mode, the roundtable mechanism mainly serves as a "pre-meeting briefing" tool. The system only performs the first round of cross-validation, generating preliminary viewpoints and evidence for each expert agent and presenting them to the human experts. Subsequent cross-examination and convergence processes are completed by the human experts through on-site discussions, while the system is responsible for recording, organizing, and assisting in the drafting. The final conclusion is based on the human experts' findings, making this mode suitable for review scenarios requiring in-depth professional judgment or strategic decision-making.

[0103] Please refer to Figure 2 The second aspect of the present invention provides a proposal review system based on evidence verification and multi-agent collaboration, comprising: The proposal preprocessing module 100 is used to perform the following steps: obtain the proposal file, and perform field validation and data cleaning on the proposal file to obtain the proposal dataset; perform semantic parsing and semantic extraction on the proposal dataset to obtain the key point vector and keyword set; The agent matching module 200 is used to match several target agents in a preset agent library based on the key point vector and the keyword set, and then link each target agent to the corresponding review database. The agent initial evaluation module 300 is used to input the key point vector and the keyword set to each target agent respectively, so that each target agent outputs the corresponding basic evaluation text according to the corresponding evaluation database, thereby obtaining several kinds of basic evaluation texts; The multi-agent cross-validation module 400 is used to perform the following steps: A review information pool is constructed based on several types of basic review texts; several target agents are cross-validated several times according to the review information pool until any round of cross-validation meets a preset convergence condition, and the review information pool is updated based on the result of that round of cross-validation. The review report generation module 500 is used to generate a structured review report based on the review information pool.

[0104] The proposal review method and system based on evidence verification and multi-agent collaboration provided by the present invention have at least the following advantages compared with the prior art: At the technical level, this invention constructs a multi-agent "roundtable" mechanism through the dynamic combination of proposal semantic understanding and multi-agent cross-validation, thereby achieving in-depth and reliable automated evaluation and enabling the review process to have the characteristics of digitalization, structure and visual traceability. At the same time, it provides a multi-industry scalable agent construction and scheduling framework.

[0105] From an economic perspective, this invention significantly reduces the time cost of reviewing proposals in the real economy and constructs a replicable proposal review system paradigm based on evidence verification and multi-agent collaboration.

[0106] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0107] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; however, any combination of these technical features that does not contradict each other should be considered within the scope of this specification.

[0108] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the concept of this application, and these improvements and substitutions should also be considered within the scope of protection of this invention. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A proposal review method based on evidence verification and multi-agent collaboration, characterized in that, include: Obtain the proposal file and perform field validation and data cleaning on the proposal file to obtain the proposal dataset; Semantic parsing and semantic extraction are performed on the proposed dataset to obtain key point vectors and keyword sets; Based on the key point vector and the keyword set, several target intelligent agents are matched in the preset intelligent agent library, and then each target intelligent agent is linked to the corresponding review database. For any of the target intelligent agents: The target intelligent agent is input with the key point vector and the keyword set, so that the target intelligent agent can retrieve a number of candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence; Based on the verification score of each candidate review evidence, a number of basic review evidence that meet the preset verification score threshold are selected from the number of candidate review evidence. The target intelligent agent generates basic review information based on the key point vector, the keyword set, and several basic review evidences, and uses the several basic review evidences and the basic review information as a set of basic review texts for the target intelligent agent; Generate several types of basic review texts based on several types of target intelligent agents; A review information pool is constructed based on several types of basic review texts; Several target intelligent agents are cross-validated several times according to the review information pool until any cross-validation in any round satisfies the preset convergence condition. The review information pool is then updated based on the result of the cross-validation in that round. A structured review report is generated based on the review information pool.

2. The proposal review method based on evidence verification and multi-agent collaboration according to claim 1, characterized in that, The process of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The intelligent agent library includes several pre-trained candidate intelligent agents, including a legal interpretation intelligent agent, a rule verification intelligent agent, a case retrieval intelligent agent, and a contingency plan retrieval intelligent agent; Among them, the review database corresponding to the legal interpretation intelligent agent is a legal clause database, which stores legal clause data; The review database corresponding to the rule verification intelligent agent is a compliance list template database, which stores industry rule text data. The review database corresponding to the case retrieval agent is a historical case database, which stores judicial judgment case document data. The review database corresponding to the plan retrieval agent is a plan template database, which stores executable proposal file data.

3. The proposal review method based on evidence verification and multi-agent collaboration according to claim 1, characterized in that, The process of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The agent library includes several pre-trained candidate agents. For any candidate agent: Obtain the review responsibility text data and review domain feature tags of the candidate agent, and then construct agent metadata based on the review responsibility text data and the review domain feature tags; Calculate the text matching score and domain matching score between the key point vector, the keyword set, and the agent metadata, and then obtain the comprehensive matching score corresponding to the candidate agent based on the text matching score and the domain matching score; Obtain the comprehensive matching score corresponding to each candidate agent, and then select the candidate agents whose comprehensive matching scores meet the preset score conditions as the target agents.

4. The proposal review method based on evidence verification and multi-agent collaboration according to claim 1, characterized in that, The step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of cross-validation, for any target agent: Based on the review information pool, the target agent is input with the basic review text corresponding to any other target agent, so that the target agent can retrieve several candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence. Based on the verification score of each candidate review evidence, several second review evidences that meet the preset verification score threshold are selected from the several candidate review evidences. The target intelligent agent generates second review information based on the basic review text and several pieces of second review evidence, and uses the several pieces of second review evidence and the second review information as a set of second review texts corresponding to the basic review text; The basic review texts corresponding to the other target intelligent agents are analyzed respectively to obtain several second review texts corresponding to the basic review texts; The review information pool is updated based on several of the second review texts.

5. The proposal review method based on evidence verification and multi-agent collaboration according to claim 4, characterized in that, The step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of the cross-validation described above: Semantic similarity clustering is performed on several second review texts to obtain several review text sets; For any set of review texts, calculate the mutual exclusion degree of review positions and the mutual exclusion degree of cited data for the set of review texts, and then obtain the degree of disagreement of review texts based on the mutual exclusion degree of review positions and the mutual exclusion degree of cited data, thereby obtaining the degree of disagreement of review texts corresponding to each set of review texts. If the divergence of all the review texts meets the preset divergence threshold, then the cross-validation round is determined to meet the convergence condition.

6. A proposal review system based on evidence verification and multi-agent collaboration, characterized in that, include: The proposal preprocessing module is used to perform the following steps: obtain the proposal file, and perform field validation and data cleaning on the proposal file to obtain the proposal dataset; Semantic parsing and semantic extraction are performed on the proposed dataset to obtain key point vectors and keyword sets; The agent matching module is used to match several target agents in a preset agent library based on the key point vector and the keyword set, and then link each target agent to the corresponding review database. The agent initial evaluation module is used to perform the following steps: For any of the target intelligent agents: The target intelligent agent is input with the key point vector and the keyword set, so that the target intelligent agent can retrieve a number of candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence; Based on the verification score of each candidate review evidence, a number of basic review evidence that meet the preset verification score threshold are selected from the number of candidate review evidence. The target intelligent agent generates basic review information based on the key point vector, the keyword set, and several basic review evidences, and uses the several basic review evidences and the basic review information as a set of basic review texts for the target intelligent agent; Generate several types of basic review texts based on several types of target intelligent agents; The multi-agent cross-validation module is used to perform the following steps: A review information pool is constructed based on several types of basic review texts; several target agents are cross-validated several times according to the review information pool until any round of cross-validation meets a preset convergence condition, and the review information pool is updated based on the result of that round of cross-validation. The review report generation module is used to generate structured review reports based on the review information pool.

7. A proposal review system based on evidence verification and multi-agent collaboration according to claim 6, characterized in that, The process of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The intelligent agent library includes several pre-trained candidate intelligent agents, including a legal interpretation intelligent agent, a rule verification intelligent agent, a case retrieval intelligent agent, and a contingency plan retrieval intelligent agent; Among them, the review database corresponding to the legal interpretation intelligent agent is a legal clause database, which stores legal clause data; The review database corresponding to the rule verification intelligent agent is a compliance list template database, which stores industry rule text data. The review database corresponding to the case retrieval agent is a historical case database, which stores judicial judgment case document data. The review database corresponding to the plan retrieval agent is a plan template database, which stores executable proposal file data.

8. A proposal review system based on evidence verification and multi-agent collaboration as described in claim 6, characterized in that, The process of matching several target agents in a preset agent library based on the key point vector and the keyword set, and then linking each target agent to a corresponding review database, includes: The agent library includes several pre-trained candidate agents. For any candidate agent: Obtain the review responsibility text data and review domain feature tags of the candidate agent, and then construct agent metadata based on the review responsibility text data and the review domain feature tags; Calculate the text matching score and domain matching score between the key point vector, the keyword set, and the agent metadata, and then obtain the comprehensive matching score corresponding to the candidate agent based on the text matching score and the domain matching score; Obtain the comprehensive matching score corresponding to each candidate agent, and then select the candidate agents whose comprehensive matching scores meet the preset score conditions as the target agents.

9. A proposal review system based on evidence verification and multi-agent collaboration as described in claim 6, characterized in that, The step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of cross-validation, for any target agent: Based on the review information pool, the target agent is input with the basic review text corresponding to any other target agent, so that the target agent can retrieve several candidate review evidences in the corresponding review database, and then perform a weighted score on each candidate review evidence to obtain a verification score for each candidate review evidence. Based on the verification score of each candidate review evidence, several second review evidences that meet the preset verification score threshold are selected from the several candidate review evidences. The target intelligent agent generates second review information based on the basic review text and several pieces of second review evidence, and uses the several pieces of second review evidence and the second review information as a set of second review texts corresponding to the basic review text; The basic review texts corresponding to the other target intelligent agents are analyzed respectively to obtain several second review texts corresponding to the basic review texts; The review information pool is updated based on several of the second review texts.

10. A proposal review system based on evidence verification and multi-agent collaboration according to claim 9, characterized in that, The step of having several target agents perform multiple rounds of cross-validation based on the review information pool until any round of cross-validation satisfies a preset convergence condition, and then updating the review information pool based on the result of that round of cross-validation, includes: In any round of the cross-validation described above: Semantic similarity clustering is performed on several second review texts to obtain several review text sets; For any set of review texts, calculate the mutual exclusion degree of review positions and the mutual exclusion degree of cited data for the set of review texts, and then obtain the degree of disagreement of review texts based on the mutual exclusion degree of review positions and the mutual exclusion degree of cited data, thereby obtaining the degree of disagreement of review texts corresponding to each set of review texts. If the divergence of all the review texts meets the preset divergence threshold, then the cross-validation round is determined to meet the convergence condition.

Images