A method for intelligent examination and generation of a science and technology project application based on multi-agent cooperation

Abstract

This invention discloses an intelligent review and generation method for science and technology project applications based on multi-agent collaboration, belonging to the field of natural language processing and intelligent auxiliary review technology. The method includes: using document structure parsing technology to parse different documents of the application into different data entities; constructing agents responsible for project funding verification, text inspection, and temporal logic verification; distributing corresponding data to the agents according to their responsibilities to obtain judgment conclusions; further determining whether there are conflicts among all judgment conclusions; using a debate mechanism to select conflicting conclusions; marking the document content corresponding to unaccepted conclusions as items to be modified; further combining the remaining conclusions with expert opinions to obtain the final generated review report; and introducing constraints to generate a project necessity analysis section. Compared with traditional methods, this invention's method possesses superior logical reasoning and conflict resolution capabilities.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent assisted review technology, and in particular to an intelligent review and generation method for science and technology project applications based on multi-agent collaboration. Technical Background In the process of applying for science and technology projects, the application materials typically include various document formats such as project summary tables, feasibility study reports, and budget tables. These documents have complex logical relationships, such as the cascading calculations and item-by-item correspondence between the budget summary table and the appendices, and the consistency of the research content hierarchy between the project summary table and the main text of the feasibility study.

[0002] Because application materials often involve multiple collaborations and rounds of revisions during the preparation and finalization stages, inconsistencies in the interpretation of the same facts are common. Manual review is not only labor-intensive but also prone to omissions, misjudgments, inconsistent standards, and difficulties in establishing a closed-loop traceability system. Traditional automated review tools are mostly based on regular expressions or static rule engines, which can handle simple format and numerical validations but struggle to handle complex cross-document logical validations and semantic compliance checks.

[0003] In recent years, with the development of large language models, censorship tools based on monolithic large models have emerged. However, these tools are prone to producing illusions when processing long texts and cannot strike a balance between rigid rules (such as strict financial rules) and flexible creative processes (such as text generation). Furthermore, monolithic models lack credibility and struggle to handle conflicts arising from multiple sources of information.

[0004] Therefore, there is an urgent need for an intelligent review system that can simulate collaboration among experts from multiple fields and possesses logical reasoning and conflict resolution capabilities. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention provides the following technical solution: S1. For the project summary, feasibility study report and funding appendix entered by the user, document structure parsing technology is used to parse key-value pairs, global context data packets and two-dimensional matrix data respectively; the global context data packet includes the main text, project budget and project time schedule; the key-value pairs include funding indicators, research content and project start and end time. S2 calls the large language model and combines the constraints and the main text of the global context data package to generate the project necessity analysis section; S3, input the project budget portion of the key-value pairs, two-dimensional matrix data and global context data into the first agent, input the main text portion of the key-value pairs and global context data into the second agent, input the project schedule portion of the key-value pairs and global context data into the third agent; S4, the first, second, and third intelligent agents perform project funding verification, text inspection, and temporal logic verification based on their respective input data, and generate corresponding conclusions based on the judgment results; S5. Use a key-value matching algorithm to identify whether there is a conflict between all conclusions. If there is no conflict, the judgment ends. If there is a conflict, the two conflicting conclusions are combined into a conflicting conclusion pair. The final accepted conclusion is determined through a debate mechanism. The document content corresponding to the unaccepted conclusion is marked as an item to be modified and the judgment ends. S6. Incorporate existing expert opinions, combine the items to be modified and all remaining conclusions to generate a review report, and output it together with the section on the necessity of the project.

[0006] Furthermore, S1 specifically refers to: For the project summary table, regular expression-driven matching is used to parse the project summary table into key-value pairs; For the expense appendix, the table parsing algorithm is used to reconstruct the expense appendix into a two-dimensional matrix data containing the subject name, value, expenditure unit and remarks; For the feasibility study report, the report is divided into independent text block indexes according to the outline level, and a global context data package is created.

[0007] Furthermore, in S2, the constraints are as follows: the writing of the project necessity analysis chapter must meet the paragraph structure of the first paragraph being the current situation, the second paragraph being the problem, the third paragraph being the countermeasure, and the fourth paragraph being the result; the second paragraph must begin with a specific sentence pattern.

[0008] Furthermore, in S3, the first intelligent agent is a combination of a large language model and a Python code interpreter, the second intelligent agent is a combination of a large language model and a text embedding model, and the third intelligent agent is a combination of a large language model and a time format processing tool; the large language model is Qwen3-32B or DeepSeek-R1 32b.

[0009] Furthermore, S4 specifically includes: For the first intelligent agent, firstly, the funding indicators in the two-dimensional matrix data and key-value pairs are compared with the corresponding values ​​in the project funding budget section of the global context data package, and the corresponding conclusion is generated based on the comparison results; secondly, the self-consistency of the funding structure in the project funding budget section of the global context data package is checked, and the corresponding conclusion is generated based on the check results. For the second agent, based on the key-value pairs and the main text of the global context data packet, according to the user-preset prohibited word library and keyword set, it sequentially determines whether there are prohibited words and confidential units in the key-value pairs and the main text of the global context data packet, and generates corresponding conclusions based on the determination results; finally, it performs a content consistency comparison between the key-value pairs and the main text of the global context data packet, and generates corresponding conclusions based on the comparison results. For the third agent, firstly, all time entities in the project time schedule part of the key-value pairs and global context data packets are extracted and standardized into a timestamp format; based on the standardized key-value pairs and the standardized global context data packet project time schedule part, time consistency checks and time logic constraint checks are performed in sequence, and corresponding conclusions are generated based on the check results.

[0010] Furthermore, the self-consistency of the funding structure in the project budget section of the global context data packet is as follows: the total project cost in the project budget section of the global context data packet equals the sum of the costs of all sub-items.

[0011] Furthermore, the specific determination process for the second intelligent agent is as follows: First, based on the prohibited word library, keyword matching is used to determine whether prohibited words exist in the key-value pairs and the main text of the global context data package. If no prohibited words exist, the corresponding conclusion is generated and the process proceeds to the next step. If prohibited words exist, a large language model is used to analyze whether there are exempt words before and after the prohibited words. Based on the analysis results, the corresponding conclusion is generated and the process proceeds to the next step. Secondly, based on the banned word list, scan the key-value pairs and the main text of the global context data to see if any confidential units appear, generate the corresponding conclusions, and proceed to the next step; Finally, using a text embedding model, the cosine similarity between the research content of the key-value pair and the main text of the global context data packet is calculated. If the similarity is lower than the threshold, the corresponding conclusion is generated and the judgment ends. If the similarity is higher than the threshold, the regular expression matching results between the key-value pair and the main text of the global context data packet and the keyword set are calculated respectively. The corresponding conclusion is generated based on the regular expression matching results and the judgment ends.

[0012] Furthermore, the time consistency check and time logic constraint check are as follows: The time consistency check verifies whether the start and end times of the items in the standardized key-value pairs are strictly consistent with the start and end times of the items in the project time schedule section of the standardized global context data package, and generates corresponding conclusions based on the check results. The time logic constraint verification is based on the standardized global context data package project schedule section. It checks whether the duration of each sub-phase in the project schedule section exceeds three months, whether the start and end times of all sub-phases fall within the project start and end time, and whether there are any cases where the start time of the later sub-phase is earlier than the end time of the previous sub-phase. The corresponding conclusions are generated based on the verification results.

[0013] Furthermore, in S5, the debate mechanism specifically involves: the large language model guiding the two agents involved in the conflicting conclusions to conduct a structured debate based on common prompt words; and further adjudicating the debate based on preset data source credibility priority rules to determine which of the conflicting conclusions is the final accepted conclusion and which is the unaccepted conclusion.

[0014] The beneficial effects of this invention are: (1) Multi-agent collaborative review architecture: Unlike traditional single large model or rule engine methods, this invention constructs a collaborative system composed of professional "expert agents" such as financial audit, semantic compliance, and temporal logic. Through role division, parallel processing and information interaction, it effectively simulates the collaboration mechanism of real review expert teams, significantly improving the professionalism, comprehensiveness and reliability of the review.

[0015] (2) Achieving the organic integration of rigid rules and flexible semantics: The system combines the semantic understanding capabilities of the large language model with external tools. While ensuring the accurate execution of rigid rules such as financial calculation and time logic, it supports in-depth semantic compliance judgment of text content, solving the problem that existing technologies cannot balance "rule rigor" and "expression flexibility".

[0016] (3) Introducing debate and conflict resolution mechanisms to ensure consistency of conclusions: In response to the typical problem of inconsistent information across documents or conflicting judgments of agents, this invention designs an arbitration process based on the blackboard model and structured debate. Through evidence backtracking, credibility priority adjudication and other methods, contradictions are automatically resolved, ensuring that the final review conclusion is unique, interpretable and traceable, and greatly reducing the cost of manual review.

[0017] (4) Supports end-to-end closed-loop generation and feedback: The system not only has intelligent review capabilities, but also automatically generates compliant project necessity chapters based on the review results, and provides modification suggestions with revision marks and structured PDF reports, realizing full-process automation from "error checking" to "error correction" to "documentation", significantly improving the efficiency and quality of application material preparation.

[0018] In summary, this invention overcomes the bottlenecks of existing science and technology project application assistance tools in cross-document logic verification, multi-source information fusion, and expert knowledge collaboration, providing a new generation of highly reliable, efficient, and available solutions for intelligent review of science and technology projects. Attached Figure Description

[0019] Figure 1 This is an overall framework diagram of a technology project application intelligent review and generation system based on multi-agent collaboration; Figure 2 This is a flowchart illustrating the workflow of the expert intelligent agent cluster module; Figure 3 This is a flowchart illustrating the workflow of the debate and conflict resolution module. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0021] The present invention proposes an intelligent review and generation system for science and technology project applications based on multi-agent collaboration. By constructing an "expert intelligent agent cluster" with different professional skills, it simulates the division of labor, review and adjudication process in real review scenarios in a multi-role collaborative manner, thereby ensuring compliance and rigor while taking into account the efficiency and quality of material generation.

[0022] like Figure 1 As shown, the method of the present invention comprises four core modules: a document parsing and task scheduling module, an expert agent cluster module, a debate and conflict resolution module, and a generation and feedback module.

[0023] The details of each module are as follows: I. Document Parsing and Task Scheduling Module This module, acting as the system's "perception center," is responsible for transforming unstructured original application documents into structured intermediate data that the intelligent agent can understand. It also generates a dynamic review task chain based on data characteristics to ensure that the intelligent agent can accurately obtain the required information in subsequent steps and rationally arrange different content and tasks according to the logical relationships between documents, thus assigning different professional intelligent agents to different tasks.

[0024] 1.1 Multimodal Document Parsing Unit The input consists of application materials uploaded by the user in Word, Excel, or PDF format, including a project summary, feasibility study report, and funding appendix. Document structure parsing technology is used to perform structural transformation based on the corresponding characteristics of the materials, specifically including: The "Project Summary" is parsed into key-value pairs using regular expression-driven matching.

[0025] For the tabular data in the funding appendix, a deep learning-based table parsing algorithm is used to identify the cell row and column span and the semantics of the table header, thereby reconstructing the funding appendix into a two-dimensional matrix data containing subject name, value, expenditure unit and remarks. For long text data in feasibility study reports, the feasibility study report is divided into independent text block indexes according to the document outline level, and a global context data package is established.

[0026] 1.2 Task Broadcasting Unit This unit instantiates a specific review task graph based on a pre-defined directed acyclic graph task dependency template and the parsed data type. This unit performs two mechanisms: static task dispatch and dynamic dependency injection.

[0027] Regarding static task distribution, firstly, the parsed project summary is distributed to each professional agent in the expert agent cluster module. Secondly, the parsed funding appendix and specific data information in the feasibility study report are distributed to different professional agents in the expert agent cluster module: the funding table data and the project funding budget-related sections of the feasibility study report are distributed to the input queue of the financial audit agent; the complete feasibility study report text is distributed to the input queue of the semantic compliance agent; and the project timeline-related sections of the feasibility study report are distributed to the input queue of the temporal logic agent.

[0028] Regarding dynamic dependency injection, this unit dynamically adjusts the review tasks based on the actual content of the funding appendix documents in the application materials. For example, when an outsourcing fee appendix is ​​identified, the system automatically adds an outsourcing fee consistency comparison task to the financial audit agent and sets the prerequisites for this task to be the completion of budget calculation and verification and outsourcing fee appendix data extraction, ensuring that the relevant data is ready before the comparison operation is performed, thereby achieving efficient task scheduling.

[0029] II. Expert Intelligent Agent Cluster Module like Figure 2 As shown, this module consists of multiple dedicated agents built on a large language model (in this specific embodiment, the large language model is Qwen3-32B or DeepSeek-R1 32b). Each agent is configured with independent system prompts, a domain knowledge base, and external tools, simulating expert roles in different domains to perform parallel reviews. The agents communicate via a message bus, enabling them to independently complete their respective review tasks and exchange information for collaborative judgment when necessary. Each agent unit ultimately generates conclusions with varying levels of detail based on user requirements.

[0030] 2.1 Financial Audit Intelligent Agent Unit The input consists of a structured project summary, funding table data, and text paragraphs from the feasibility study report regarding the project budget. This agent unit is equipped with a Python code interpreter, responsible for all numerical calculations, verifications, and logical inferences.

[0031] The specific working mechanism is as follows: The agent first scans all input numerical fields and cleans the format. Then, it converts the expense calculation rules described by humans in natural language into a Python verification script and strictly performs precision control calculations, including keeping one decimal place and rounding.

[0032] Secondly, the intelligent agent is responsible for performing cross-table linkage checks, reading key funding indicators such as total cost and labor cost fields from the project summary table, and comparing them with the corresponding values ​​in the project budget section of the feasibility study report and the corresponding values ​​in the funding appendix. During the comparison process, it not only checks whether the values ​​themselves are consistent, but also verifies whether the units of the values ​​are consistent, such as whether they are all in ten thousand yuan.

[0033] Furthermore, the AI ​​can perform self-consistency verification of the internal funding structure of the feasibility study report, and logically verify that the total project cost equals the sum of all sub-items. It automatically extracts the amounts of each sub-item such as direct costs and labor costs, performs a summation operation, and compares it with the total cost field. If the results are consistent, it is considered to have passed the logical verification that the total project cost equals the sum of all sub-items.

[0034] In addition, for tasks distributed in dynamic dependency injection, each cost item in the feasibility study report must be checked for consistency with the costs in the outsourcing fee appendix. Any inconsistencies should be immediately flagged as numerical discrepancies.

[0035] Finally, the financial auditing intelligent agent unit generates a financial audit conclusion based on the judgment results of all the above aspects.

[0036] 2.2 Semantic Compliant Intelligent Agent Unit The input consists of a project summary, the main text of the feasibility study report, and a pre-defined knowledge base of prohibited words. This intelligent agent unit is equipped with a text embedding model and is responsible for performing tasks such as semantic understanding of the text content, terminology standardization checks, and cross-document content consistency verification.

[0037] In terms of prohibited word scanning, this unit identifies key target words through keyword matching, and uses a large model to analyze the word collocation and context before and after them, and determines whether there are exempt words such as "prototype" and "simulation", thereby accurately distinguishing prohibited terms. Regarding the management of partner organizations, this unit scans documents to check for any partner organization names that should not be mentioned, thereby ensuring that the use of organization information complies with regulations. In terms of content consistency comparison, this unit extracts the research content field in the project summary table and all the content under the corresponding chapter subheadings of the feasibility study report. It uses a text embedding model to calculate the cosine similarity between the two. If the similarity is lower than the preset threshold, it is marked as inconsistent in content. If the similarity is higher than the preset threshold, it calculates the regular expression matching results between the two and the preset core keywords. If the core keywords are missing, it is marked as inconsistent in content.

[0038] Finally, the semantic compliance intelligent agent unit generates a semantic compliance conclusion based on the judgment results of all the above aspects.

[0039] 2.3 Sequential Logic Intelligent Agent Unit The input consists of the project summary and the schedule section of the feasibility study report. This intelligent agent unit is responsible for constructing the complete project timeline and deriving its logical rationality. Its core task is to ensure that the project's time plan remains consistent across multiple documents and that the time arrangements for each stage comply with objective laws and the constraints of the application specifications.

[0040] First, the agent extracts all time entities from the project summary and the schedule section of the feasibility study report, including the overall start and end times of the project, the start and end times of each stage, etc., and uses time parsing algorithms (such as Chrono) to standardize various time expressions (such as January 2024, January 2024, 2024.01, etc.) into a timestamp format for easy subsequent calculation and comparison.

[0041] Secondly, multi-source time consistency verification was conducted. The agent verified whether the project start and end times described in the schedule section of the feasibility study report were strictly consistent with the start and end times filled in the project summary, ensuring that the time information in the two core documents was completely matched.

[0042] Furthermore, time logic constraint verification is carried out. The agent calculates whether the duration of each sub-phase exceeds three months, checks whether all activity time points fall within the total project cycle, and detects whether there is a logical paradox where the start time of the later phase is earlier than the end time of the previous phase, to ensure that the project schedule conforms to the laws of the physical world and the application requirements.

[0043] Finally, the temporal logic agent unit generates a temporal logic conclusion based on the judgment results of all the above aspects.

[0044] III. Debate and Conflict Resolution Module In complex application review scenarios, the same fact may appear in different forms in documents, and different agents may reach inconsistent conclusions based on their respective professional perspectives. In such cases, a fair arbitration mechanism is needed to determine the final credible conclusion. This module introduces a blackboard mode and a debate mechanism to resolve misunderstandings of the same information by different agents or conflicts in data sources, ensuring the uniqueness and accuracy of the review conclusion. The workflow is as follows: Figure 3 As shown.

[0045] 3.1 Collision Detection Unit This unit maintains a shared memory space, the global blackboard, where the outputs of all expert agents are written in structured data format. Each record on the blackboard contains key information about the conclusion, such as the name of the review metric, its value or judgment result, the data source, and the identifier of the agent that generated the conclusion.

[0046] This unit monitors the key-value pairs with the same name on the blackboard in real time. Using a key-value matching algorithm, it identifies the conclusions generated by each professional agent in the expert agent cluster module for the same review object. If inconsistencies are found in the values ​​of the same indicator, or if there are contradictory affirmative and negative conclusions for the same judgment question, the unit immediately locks the relevant entries, preventing these contradictory conclusions from flowing into the final report, and triggers a conflict resolution process, inputting the conflict event pair into the debate arbitration unit. If all conclusions are conflict-free, all review conclusions from the agent cluster are directly transmitted to the generation and feedback module.

[0047] 3.2 Arbitration Unit The debate arbitration unit receives locked conflict facts as input and organizes structured debates among specialized intelligent agents involved in the conflict.

[0048] This unit guides each agent to present their reasoning and chain of evidence using carefully designed universal prompts. For example, when a financial audit agent calculates a specific amount for an expense based on a table, while a semantic compliance agent interprets it as a different amount based on the text description, the debate and arbitration unit requires the financial agent to demonstrate the table data used, the summation formula, the calculation steps, and the final result. Simultaneously, the semantic agent is required to demonstrate the original text paragraphs from which it extracted information, the text parsing logic, and the basis for its semantic understanding. Subsequently, the unit introduces a metacognitive arbitration model with built-in data source credibility priority rules. Based on preset rules (such as calculated values ​​being superior to text description values), the model adjudicates the evidence from both sides, establishes the final facts, and marks the corresponding document content of the losing party as items requiring modification.

[0049] IV. Generation and Feedback Module This module performs specific text generation and revision operations based on the parsed application materials and review results. Specifically, it generates a necessity section based on the current situation analysis and significance of the results provided in the feasibility study report, and modifies the text based on expert opinions, thus achieving a closed loop from error checking to error correction.

[0050] 4.1 Structured Content Generation Unit The input consists of the current situation analysis, the significance of the results, and relevant constraints from the feasibility study report. This unit utilizes a large language model and applies a "current situation-problem-countermeasure-result" thought chain template to automatically generate a project necessity analysis that meets word count limits and paragraph structure requirements.

[0051] During the generation process, the unit performs dynamic constraint injection, which forces the second paragraph to start with a specific sentence structure. Immediately after generation, a lightweight validator (i.e., a lightweight large language model) is called to check whether the constraints are met. If not, a rewriting mechanism is automatically triggered until the content is fully compliant.

[0052] 4.2 Intelligent Revision and Report Rendering Unit The input consists of expert review comments and the review conclusions of the intelligent agent cluster. The expert review comments are obtained by inviting multiple experts to review the application documents.

[0053] In response to expert opinions, this unit uses a semantic similarity retrieval algorithm to locate relevant paragraphs in the original text, generates revised text based on a large language model, and retains revision pattern markers for user comparison. Finally, the unit aggregates all warnings, errors, and revision suggestions output by all agents, categorizes them by severity, and generates a structured PDF review report with jump links and detailed revision suggestions, visually presenting the review results.

[0054] Based on the aforementioned core modules, this invention proposes an intelligent review and generation method for science and technology project applications based on multi-agent collaboration. The overall process is as follows: 1. Users upload a project summary, feasibility study report, and budget appendix.

[0055] 2. The multimodal document parsing unit in the document parsing and task scheduling module performs multimodal parsing on the project summary, feasibility study report and appendices uploaded by the user, transforming the unstructured documents into intermediate data readable by the intelligent agent, and transmitting it to the review task broadcasting unit; the review task broadcasting unit transmits different data to different professional intelligent agents in the expert intelligent agent cluster module according to the data attributes.

[0056] 3. All specialized agents in the expert agent cluster module receive data, begin working in parallel, generate corresponding conclusions, and transmit them to the debate and conflict resolution module. Among them, the "financial audit agent" uses a toolchain to perform accurate calculations and logical verification of funding figures, the "semantic compliance agent" performs compliance checks on text content based on a knowledge base, and the "temporal logic agent" is responsible for the reasonable deduction of project progress.

[0057] 4. The debate and conflict resolution module receives the conclusions from all professional agents, and the conflict monitoring unit determines whether there are conflicts between the conclusions. If there are conflicts in the conclusions reached by the agents (such as inconsistencies between the simplified table amount and the main text description), the debate arbitration unit intervenes and determines the final accepted conclusion through the debate mechanism, marking the unaccepted conclusions as items to be modified; if there are no conflicts, the conclusions generated by all agents are directly transmitted to the generation and feedback module.

[0058] 5. The generation and feedback module receives the items to be modified, the agent's conclusions, and the parsed application documents from the debate and conflict resolution module. Based on expert opinions and review results, it automatically generates a review report and calls upon the large language model to generate a project necessity analysis section according to the constraints. Finally, a complete review report and project necessity analysis section are obtained.

[0059] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for intelligent review and generation of a science and technology project application based on multi-agent collaboration, characterized in that, include: S1. For the project summary, feasibility study report and funding appendix entered by the user, document structure parsing technology is used to parse key-value pairs, global context data packets and two-dimensional matrix data respectively; the global context data packet includes the main text, project budget and project time schedule; the key-value pairs include funding indicators, research content and project start and end time. S2 calls the large language model and combines the constraints and the main text of the global context data package to generate the project necessity analysis section; S3, input the project budget portion of the key-value pairs, two-dimensional matrix data and global context data into the first agent, input the main text portion of the key-value pairs and global context data into the second agent, input the project schedule portion of the key-value pairs and global context data into the third agent; S4, the first, second, and third intelligent agents perform project funding verification, text inspection, and temporal logic verification based on their respective input data, and generate corresponding conclusions based on the judgment results; S5. Use a key-value matching algorithm to identify whether there is a conflict between all conclusions. If there is no conflict, the judgment ends. If there is a conflict, the two conflicting conclusions are combined into a conflicting conclusion pair. The final accepted conclusion is determined through a debate mechanism. The document content corresponding to the unaccepted conclusion is marked as an item to be modified and the judgment ends. S6. Incorporate existing expert opinions, combine the items to be modified and all remaining conclusions to generate a review report, and output it together with the section on the necessity of the project. 2.The method of claim 1, wherein, Specifically, S1 is: For the project summary table, regular expression-driven matching is used to parse the project summary table into key-value pairs; For the expense appendix, the table parsing algorithm is used to reconstruct the expense appendix into a two-dimensional matrix data containing the subject name, value, expenditure unit and remarks; For the feasibility study report, the report is divided into independent text block indexes according to the outline level, and a global context data package is created.

3. The intelligent review and generation method for science and technology project applications based on multi-agent collaboration as described in claim 1, characterized in that, In S2, the constraints are as follows: the writing of the project necessity analysis chapter must meet the paragraph structure of the first paragraph being the current situation, the second paragraph being the problem, the third paragraph being the countermeasure, and the fourth paragraph being the result; the second paragraph must begin with a specific sentence pattern.

4. The method for intelligent review and generation of science and technology project applications based on multi-agent collaboration as described in claim 1, characterized in that, In S3, the first intelligent agent is a combination of a large language model and a Python code interpreter, the second intelligent agent is a combination of a large language model and a text embedding model, and the third intelligent agent is a combination of a large language model and a time format processing tool; the large language model is Qwen3-32B or DeepSeek-R1 32b.

5. The intelligent review and generation method for science and technology project applications based on multi-agent collaboration as described in claim 1, characterized in that, Specifically, S4 is: For the first intelligent agent, firstly, the funding indicators in the two-dimensional matrix data and key-value pairs are compared with the corresponding values ​​in the project funding budget section of the global context data packet, and the corresponding conclusions are generated based on the comparison results. Secondly, verify the consistency of the funding structure in the project funding budget section of the global context data package, and generate corresponding conclusions based on the verification results; For the second agent, based on the key-value pairs and the main text of the global context data packet, according to the user-preset prohibited word library and keyword set, it sequentially determines whether there are prohibited words and confidential units in the key-value pairs and the main text of the global context data packet, and generates corresponding conclusions based on the determination results; finally, it performs a content consistency comparison between the key-value pairs and the main text of the global context data packet, and generates corresponding conclusions based on the comparison results. For the third agent, firstly, all time entities in the project time schedule part of the key-value pairs and global context data packets are extracted and standardized into a timestamp format; based on the standardized key-value pairs and the standardized global context data packet project time schedule part, time consistency checks and time logic constraint checks are performed in sequence, and corresponding conclusions are generated based on the check results.

6. The method for intelligent review and generation of science and technology project applications based on multi-agent collaboration as described in claim 5, characterized in that, The self-consistency of the funding structure in the project budget section of the global context data packet is as follows: the total project cost in the project budget section of the global context data packet is equal to the sum of the costs of all sub-items.

7. The intelligent review and generation method for science and technology project applications based on multi-agent collaboration as described in claim 5, characterized in that, The specific determination process for the second intelligent agent is as follows: First, based on the prohibited word library, keyword matching is used to determine whether there are prohibited words in the key-value pairs and the main text of the global context data packet. If no prohibited words are found, the corresponding conclusion is generated and the process proceeds to the next step. If a prohibited word exists, the large language model is used to analyze whether there are exempt words before and after the prohibited word. Based on the analysis results, the corresponding conclusion is generated and the process proceeds to the next step. Secondly, based on the banned word list, scan the key-value pairs and the main text of the global context data to see if any confidential units appear, generate the corresponding conclusions, and proceed to the next step; Finally, using a text embedding model, the cosine similarity between the research content of the key-value pair and the main text of the global context data packet is calculated. If the similarity is lower than the threshold, the corresponding conclusion is generated and the judgment ends. If the similarity is higher than the threshold, the regular expression matching results between the key-value pair and the main text of the global context data packet and the keyword set are calculated respectively. The corresponding conclusion is generated based on the regular expression matching results and the judgment ends.

8. The intelligent review and generation method for science and technology project applications based on multi-agent collaboration as described in claim 5, characterized in that, The time consistency check and time logic constraint check are as follows: The time consistency check verifies whether the start and end times of the items in the standardized key-value pairs are strictly consistent with the start and end times of the items in the project time schedule section of the standardized global context data package, and generates corresponding conclusions based on the check results. The time logic constraint verification is based on the standardized global context data package project schedule section. It checks whether the duration of each sub-phase in the project schedule section exceeds three months, whether the start and end times of all sub-phases fall within the project start and end time, and whether there are any cases where the start time of the later sub-phase is earlier than the end time of the previous sub-phase. The corresponding conclusions are generated based on the verification results.

9. The method for intelligent review and generation of science and technology project applications based on multi-agent collaboration as described in claim 1, characterized in that, In S5, the debate mechanism specifically involves: the large language model guiding the two agents involved in the conflict conclusions to conduct a structured debate based on common prompt words; and further adjudicating the debate based on preset data source credibility priority rules to determine which of the conflict conclusions is the final accepted conclusion and which is the unaccepted conclusion.

Images