A project proposal review assistance method based on a large language model

Abstract

The application relates to the technical field of artificial intelligence and document automatic processing, in particular to an auxiliary project establishment auditing method based on a large language model, which constructs an auxiliary project establishment auditing system based on artificial intelligence, extracts core auditing dimension content from project declaration materials, mines a technical keyword sequence of the project declaration materials based on the core auditing dimension content, retrieves a plurality of technical comparison documents from a technical document library according to the technical keyword sequence, and then automatically generates an AI auditing evaluation report of the project declaration materials, fundamentally solves the technical problems of high labor cost, difficult experience reuse, strong one-sidedness of evaluation and review, and large decision risk in traditional project establishment auditing, realizes efficient, standardized and process-oriented project establishment auditing, makes the auditing process more objective and rigorous, and the auditing result more reliable, and simultaneously improves the traceability of the auditing conclusion and the standardization of the report format.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and document automation processing technology, specifically to an auxiliary project approval method based on a large language model. Background Technology

[0002] The review and approval of scientific research projects is a key link in scientific research management and the allocation of scientific and technological resources. Its role is to comprehensively evaluate the innovation, feasibility, technical level and application value of the proposed projects, and to provide a basis for decision-making on whether to approve the projects.

[0003] Currently, project approval generally adopts a manual meeting review model: the reviewing unit collects project materials, organizes experts in the field for centralized review, and draws a conclusion after presentations, Q&A sessions, scoring, and summarization. However, this model relies heavily on expert experience and subjective judgment, and has many technical shortcomings in practical application. (1) Expert experience is difficult to reuse and pass on: the professional knowledge and review experience of experts only exist in their own minds and cannot form a standardized knowledge base for subsequent reviews, resulting in the review standards fluctuating with the change of experts; (2) The problem of bias in the review is prominent: Due to the limitations of individual experts' cognition, the scope of their professional fields, and subjective preferences, the review results are prone to bias and it is difficult to comprehensively and objectively evaluate the technical value and innovation level of the project. (3) High labor costs and low efficiency: Each review requires a meeting of multiple experts, which consumes a lot of time and labor costs, and the review cycle is long, which cannot meet the needs of centralized review of large-scale projects. (4) High decision-making risk: The project approval results rely heavily on the personal opinions of experts, lack objective data support and quantitative evaluation standards, which can easily lead to decision-making errors and cause huge opportunity costs and waste of resources.

[0004] To alleviate the above problems, the industry has begun to try using information technology, automation, and even artificial intelligence to assist in project approval. However, these methods still have problems such as difficulty in effectively handling unstructured application materials, lack of authoritative patent documents as objective evidence, non-standard output of review reports, and inability to trace and verify review conclusions. The output of review results is not accurate enough, has poor reliability, and low practicality, and cannot fundamentally solve the core pain points of traditional project approval.

[0005] Therefore, how to provide a more stable, objective, standardized and practical method to assist in project approval has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing an auxiliary project approval method based on a large language model. This method involves constructing an AI-based auxiliary project approval system that extracts core approval dimensions from project application materials. Based on these core dimensions, it mines technical keyword sequences from the application materials and retrieves comparative technical literature from a technical literature database. Subsequently, it automatically generates an AI-based review and evaluation report for the project application materials. This fundamentally solves the technical problems of high labor costs, difficulty in reusing experience, strong bias in review, and high decision-making risks in traditional project approval processes. It achieves higher efficiency, standardization, and streamlined processes in project approval, making the review process more objective and rigorous, the results more reliable, and improving the traceability of review conclusions and the standardization of report formats.

[0007] The objective of this invention is achieved through the following approach: A method for assisting in project approval based on a large language model includes the following steps: 1) Construct an AI-based auxiliary project approval system, including a content extraction module, a feature mining module, a literature retrieval module, a report generation module, and a deep question answering module; 2) Use the content extraction module to extract core review dimensions from project application materials; 3) Based on the core review dimensions extracted in step 2), use the feature mining module to mine the technical keyword sequence in the project application materials; 4) The technical keyword sequence is vectorized using the literature retrieval module, and combined with the similarity algorithm, several technical comparison documents are retrieved from the technical literature database; 5) Based on the project application materials and the technical comparison literature obtained in step 4), use the report generation module to generate an AI review and evaluation report for the project application materials.

[0008] Preferably, the content extraction module is used to accurately locate and extract core review dimension information from project application materials through a dual-track mechanism combining regular expression rule matching and semantic recognition of large language models, and to construct a standardized core review corpus. The feature mining module is used to drive a large language model through prompt word engineering to extract highly discriminative technical keyword sequences from core review dimension information, providing core basis for literature retrieval. The document retrieval module is used to vectorize the sequence of technical keywords into project technical feature vectors, and to use a vector similarity algorithm to calculate the similarity between the project technical feature vectors and various technical documents in the technical document database. After multi-dimensional weighted sorting, several technical comparison documents are selected to construct a dynamic knowledge source for review. The report generation module is used to generate an in-depth review and evaluation draft of the project application materials based on technical comparison literature and project application materials through search-enhanced generation technology, and then format and package it into a standardized AI review and evaluation report. The deep question-and-answer module is used to respond to users' questions about the review report, and output professional answers with evidence citations by combining the associated knowledge base, so as to realize closed-loop verification and traceability of the review logic; The data output terminal of the content extraction module is connected to the data input terminal of the feature mining module, the data output terminal of the feature mining module is connected to the data input terminal of the document retrieval module, the data output terminal of the document retrieval module is connected to the data input terminal of the report generation module, and the data output terminal of the report generation module is connected to the data input terminal of the deep question answering module.

[0009] Preferably, in step 2), the core review dimension content is extracted in the following manner: 2-1) Based on the directory structure of the project application materials, use regular expressions to locate the target title of the project application materials, extract the text content of the target title, and determine whether the text content is complete. If it is incomplete, proceed to step 2-2). 2-2) Construct content extraction instructions, and input the project application materials and content extraction instructions into the large language model, and use the large language model to extract key paragraphs from the project application materials; 2-3) The text content extracted in step 2-1) is deduplicated and merged with the key paragraphs identified by the large language model to obtain the core review dimension content.

[0010] Preferably, in step 3), the sequence of technical keywords is obtained in the following manner: 3-1) Based on the core review dimensions extracted in step 2), use the prompt word engineering to construct technical feature extraction instructions, and drive the large language model as the technical feature extraction model; 3-2) Input the preset technical dimension guidance instructions into the technical feature extraction model, and use the technical feature extraction model to perform deep semantic analysis on the text content of the project application materials to obtain the technical keyword sequence of the project application materials.

[0011] Preferably, in step 4), the comparative technical literature is obtained in the following manner: 4-1) Map the sequence of technical keywords obtained in step 3) to a high-dimensional vector space to construct the project feature vector; 4-2) Using the preset patent knowledge base and the user-uploaded associated document knowledge base as the technical document library, the vector similarity algorithm is used to calculate the semantic correlation between the project feature vector and each technical document in the technical document library; 4-3) Based on semantic relevance, select multiple technical documents as review reference sources from high to low; 4-4) Based on technical relevance, publication time, and legal status, the technical documents in the review reference sources are sorted by weight for a second time to select a number of technical comparison documents.

[0012] Preferably, in step 5), the AI ​​review and evaluation report is generated in the following manner: 5-1) The selected technical comparison literature and project application materials are used as context and injected into the prompts of the large language model. The large language model is then guided to generate an in-depth review and evaluation draft of the project application materials through search enhancement generation technology. 5-2) Formatting constraint instructions are constructed based on the JSON Schema formatting protocol. The formatting constraint instructions are used to drive the large language model to transform the initial draft of the in-depth review and assessment into structured data that conforms to the preset field definitions. 5-3) Using a document rendering engine, structured data is mapped to report templates, and AI audit and assessment reports in DOCX format are automatically packaged and generated.

[0013] Preferably, the formatting constraint instructions include: Pre-defined structured data description documents conforming to the JSON Schema protocol; The description document is incorporated into the prompt word instruction, constraining the large language model to output JSON format data that conforms to the field definition; The JSON formatted data is populated into a preset document template using a document rendering engine.

[0014] Preferably, it also involves a process of responding to user questions using a deep question-answering module, specifically including: (1) Perform semantic analysis on the questions entered by the user regarding the AI ​​review and evaluation report, and extract the core points of doubt; (2) Based on the core questions, re-search the text fragments in the technical comparison literature that are most relevant to the project application materials; (3) Input the user questions, project application materials and text fragments obtained in step (2) into the large language model, use the large language model to output professional answers, and mark the source information of the cited literature in the reply.

[0015] The beneficial effects of this invention include the following: ① This invention employs a dual-track extraction mechanism combining regular expression rule matching and large language model semantic recognition. First, regular expressions are used to quickly locate research content, technical routes, and innovative points based on the application document's table of contents hierarchy, and then extract text blocks, ensuring efficient and stable extraction of structured documents. For unstructured materials with irregular table of contents or incomplete extraction, a large language model is used to perform deep semantic recognition to accurately complete key information. Finally, the two types of results are deduplicated and merged to form a standardized corpus. This approach overcomes the shortcomings of single extraction methods that rely on fixed formats or insufficient understanding. The combination of hardware and software, with their complementary advantages, significantly improves the automation level and recognition accuracy of extracting core review information from massive amounts of application materials.

[0016] ② This invention ensures the authenticity and credibility of review conclusions through Search Enhancement Generation (RAG) technology: First, based on technical keywords, a hybrid search and multi-dimensional re-ranking are performed from a technical literature database to select highly relevant patent documents and related documents to constitute authoritative knowledge sources; then, the context of the documents and project materials are injected into a large language model, strictly constraining the model to generate evaluation content only based on external evidence. This mechanism of the invention can eliminate false information and logical illusions that general large models are prone to produce in professional fields from the source, ensuring that every conclusion in the report is verifiable and traceable, and improving the scientificity, objectivity, and authority of the project approval results.

[0017] ③ This invention achieves standardized report generation through JSON Schema format constraints and automated document rendering. First, it uses a preset JSON Schema to force the large language model to output structured data with stable structure and complete fields, avoiding formatting errors and logical omissions. Then, the document rendering engine automatically fills the data into a preset DOCX template, completing the unified layout of fonts, tables, headers and footers, and page layout. This completely solves the problems of inconsistent formats, chaotic content, and high review costs associated with manual writing. The generated reports can be directly integrated into office workflows, significantly reducing the workload of later sorting, proofreading, and typesetting.

[0018] ④ This invention constructs a complete review loop through knowledge-enhanced deep interactive question-and-answer. When a user raises questions about the report, the system first analyzes the question and extracts the core points of doubt, and then retrieves relevant literature to obtain evidence fragments. Subsequently, the system combines project materials and search results to drive the model to generate professional answers with citation annotations, supporting the viewing of the original text.

[0019] This invention addresses the shortcomings of traditional reports that "only provide conclusions without supporting evidence," making the review logic explainable, the conclusions traceable, and the evidence verifiable. This significantly improves the transparency and persuasiveness of the review process, providing more reliable support for project initiation decisions.

[0020] Definitions: Prompt Engineering refers to a series of methods, techniques, and practices that involve designing, optimizing, and debugging prompts input to a large language model, standardizing instructions, context, format, and examples, guiding the model to accurately understand intent, and output content, format, and logic that meet expectations, thereby efficiently utilizing the capabilities of the large model.

[0021] Large Language Model (LLM) is a large-scale parameter-based model pre-trained on massive text corpora based on deep learning and the Transformer architecture. It can understand and generate human natural language and perform general language tasks such as understanding, translation, question answering, creation, reasoning, and code generation.

[0022] JSON Schema format: "JSON Schema" is a JSON document (which must be an object). "JSON Schema format" is a metadata description format written in JSON, used to standardize the description of the structure, type, constraints and semantics of JSON data.

[0023] JSON Schema protocol: It is an open, cross-language, platform-independent standard protocol led by the IETF. It specifies the syntax, semantics, validation rules, version evolution and media types of JSON Schema, and is used for exchanging data contracts between systems, automating validation, document generation and interface interaction control. Attached Figure Description

[0024] Figure 1 This is a diagram of the overall architecture of the present invention; Figure 2 This is a flowchart illustrating the AI-based review and assessment report generation process in an embodiment of the present invention. Figure 3 This is a flowchart of the present invention. Detailed Implementation

[0025] like Figures 1 to 3 As shown, a method for assisting project approval based on a large language model includes the following steps: 1) Construct an AI-based auxiliary project approval system, including a content extraction module, a feature mining module, a literature retrieval module, a report generation module, and a deep question answering module; 2) Use the content extraction module to extract core review dimensions from project application materials; 3) Based on the core review dimensions extracted in step 2), use the feature mining module to mine the technical keyword sequence in the project application materials; 4) The technical keyword sequence is vectorized using the literature retrieval module, and combined with the similarity algorithm, several technical comparison documents are retrieved from the technical literature database; 5) Based on the project application materials and the technical comparison literature obtained in step 4), use the report generation module to generate an AI review and evaluation report for the project application materials.

[0026] The following is an example of implementing the above method: 1) Construct an AI-based auxiliary project approval system, including a content extraction module, a feature mining module, a literature retrieval module, a report generation module, and a deep question answering module; The content extraction module is used to accurately locate and extract core review dimension information such as research content, technical path, and core innovation points from unstructured project application materials in multiple formats such as PDF and Word through a dual-track mechanism that combines regular expression rule matching and semantic recognition of large language models, and to build a standardized core review corpus. The dual-track mechanism includes: Rule matching path: Based on the directory structure of the application materials, use regular expressions to locate the target title and extract its corresponding text content; Semantic recognition path: Construct content extraction instructions to drive the large language model to identify and extract preset research content paragraphs from unstructured text.

[0027] The feature mining module is used to drive a large language model through prompt word engineering to extract highly discriminative technical keyword sequences (such as technical fields, core algorithms, and application scenarios) from core review dimension information, providing core basis for literature retrieval. The document retrieval module is used to vectorize the sequence of technical keywords into project technical feature vectors, and to use a vector similarity algorithm to calculate the similarity between the project technical feature vectors and various technical documents in the technical document database. After multi-dimensional weighted sorting, several technical comparison documents are selected to construct a dynamic knowledge source for review. The report generation module is used to generate an in-depth review and evaluation draft of the project application materials based on technical comparison literature and project application materials through search-enhanced generation (RAG) technology, and then format and package it into a standardized AI review and evaluation report. The in-depth question-and-answer module is used to respond to users' questions about the audit report, and outputs professional answers with evidence citations by combining the associated knowledge base, so as to realize the closed-loop verification and traceability of the audit logic.

[0028] The data output terminal of the content extraction module (outputting core review dimension content) is connected to the data input terminal of the feature mining module. The data output terminal of the feature mining module (outputting technical keyword sequences) is connected to the data input terminal of the literature retrieval module. The data output terminal of the literature retrieval module (outputting technical comparison literature) is connected to the data input terminal of the report generation module. The data output terminal of the report generation module (outputting AI review and evaluation report) is connected to the data input terminal of the deep question answering module. The deep question answering module outputs question and answer responses with evidence.

[0029] 2) Determine the project application materials (e.g., PDF, Word format) to be reviewed and input them into the project review auxiliary system. Use the content extraction module to extract the core review dimensions from the project application materials. The specific process is as follows: In this embodiment, before extracting the core review dimensions from the project application materials, it is necessary to preprocess the project application materials, including format conversion, text cleaning and noise reduction. For scanned PDF documents, OCR technology is used for text recognition and table structure recognition.

[0030] 2-1) Set up a regular expression engine and, based on the directory structure of the project application materials, use regular expressions to locate the target title of the project application materials, extract the text content of the target title, and determine whether the text content is complete. If it is incomplete, proceed to step 2-2). For example, for the directory structure, the regular expression operator r'Research Content(.*?)\n\d+' is used to identify specific level headings in the application. If the system detects "II. Research Content" or its synonyms in the text, it automatically extracts the entire text block from below that heading to the beginning of the next heading. If the regular expression fails to obtain complete information or the directory is not standardized, proceed to step 2-2), and input the original text into the large language model.

[0031] 2-2) Construct a content extraction instruction (Prompt), such as "Please identify the description of the research technical path and core innovation points from the following documents, keep the original wording and do not embellish it", and input the project application materials and content extraction instruction into the large language model, and use the large language model to extract the key paragraphs of the project application materials; 2-3) Merge the structured text extracted by regular expressions in step 2-2) with the key paragraphs identified by the model to form a core review dimension corpus.

[0032] 3) Based on the core review dimensions extracted in step 2), the technical keyword sequence of the project application materials is mined using the feature mining module. The specific process is as follows: 3-1) Based on the core review dimensions extracted in step 2), use PromptEngineering to construct technical feature extraction instructions, driving the large language model as the technical feature extraction model. For example, if the role is set as "Senior Patent Examination Expert", the instructions include: "Analyze the following project text and extract 8-15 core technical keywords. The keywords must cover the technical field, core algorithm, physical problem solved and application scenario". 3-2) Input the preset technical dimension guidance instructions into the technical feature extraction model, and use the technical feature extraction model to perform deep semantic analysis on the text content of the project application materials, filter out redundant adjectives, and obtain a highly distinctive sequence of technical keywords, such as "heterogeneous computing", "semantic segmentation", and "dynamic load balancing".

[0033] 4) The technical keyword sequence is vectorized using the literature retrieval module, and combined with a similarity algorithm, several comparative technical documents are retrieved from the technical literature database. The specific process is as follows: 4-1) Call the pre-trained Embedding model to map the sequence of technical keywords obtained in step 3) to a high-dimensional continuous vector space to construct the project feature vector; 4-2) The pre-built global patent database (whose index contains feature vectors such as patent abstracts and claims) and the user-uploaded related document knowledge base (also called related document knowledge base) are used as the technical document base. The cosine similarity algorithm is used to calculate the cosine value of the angle between the project feature vector and each technical document in the technical document base. 4-3) Based on the cosine value of the included angle, select the top 50 patent documents and the top 50 related documents from high to low as the initial set; 4-4) Based on technical relevance, publication time, and legal status, the technical documents in the examination reference sources are sorted by weight for a second time, and finally the most relevant documents (up to 25 patents and up to 25 related documents) are selected.

[0034] 5) Based on the project application materials and the technical comparison literature obtained in step 4), the specific method for generating an AI review and evaluation report for the project application materials using the report generation module is as follows: 5-1) The abstracts of the selected technical comparison literature and the text content of the project application materials are used as context and injected into the model prompts of the large language model. The large language model is guided by the search-enhanced generation (RAG) technology to generate an in-depth review and evaluation draft containing "key indicators", "advantage indicators", "disadvantage indicators", "value score and risk expectation" and other contents. 5-2) In order to solve the technical problem of unstable output format of large language models, formatting constraint instructions are constructed based on the JSON Schema formatting protocol to drive the model to transform the initial draft of the deep review and evaluation into structured data that conforms to the preset field definitions; For example, a standard JSON schema can be predefined, defining fields such as {"title": "AI audit assessment report document template", "description": "JSON structure used to format LLM output as an audit assessment report", "type": "object", "properties": {}}. By adding this schema requirement to the Prompt, the model is forced to output JSON data according to this structure.

[0035] 5-3) Using a document rendering engine (a renderer wrapped in Python-Docx), structured data is mapped to a DOCX report template, automatically completing font settings, table drawing, and header and footer generation, and finally outputting an AI audit and evaluation report in DOCX format.

[0036] The formatting constraint instructions include: Pre-defined structured data description document conforming to the JSON Schema protocol; The description document is incorporated into the prompt word instruction, constraining the large language model to output JSON format data that conforms to the field definition; The JSON formatted data is populated into a preset document template using a document rendering engine.

[0037] After generating the report, this embodiment also provides users with a dialogue-based in-depth exploration entry point. This function is implemented by the deep question-answering module, specifically including: (1) Perform semantic analysis on the questions entered by the user regarding the AI ​​review and evaluation report, and extract the core points of doubt; (2) Based on the core questions, re-search the text fragments in the technical comparison literature that are most relevant to the project application materials; (3) Input the user questions, project application materials and text fragments obtained in step (2) into the large language model, use the large language model to output professional answers, and mark the source information of the cited literature in the reply.

[0038] For example, when a user asks a question about the report's conclusions (such as "Why do you think this project is similar to patent A?"), the system re-searches for patent excerpts in the technical comparison literature, and then uses a large model to combine the user's question, the original application text, and detailed comparison information of related patents to output a professional answer with citations (such as "[Reference 1] mentions on page 5, line 3..."), thereby constructing a traceable and auditable closed loop for review and evaluation.

[0039] In summary, this invention, through a complete technical process encompassing content extraction, feature mining, literature retrieval, report generation, and deep question answering, employs key methods such as dual-track content extraction, hybrid literature retrieval, RAG-enhanced generation, and JSON Schema format constraints. This achieves automation, standardization, and intelligence in the review and approval of research projects. It can efficiently process unstructured application materials, objectively complete technical comparisons and project evaluations, output standardized and well-supported review reports, and support conclusion tracing and interactive verification, significantly improving the overall efficiency and reliability of project approval work.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications made to the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for assisting project approval based on a large language model, characterized in that, Includes the following steps: 1) Construct an AI-based auxiliary project approval system, including a content extraction module, a feature mining module, a literature retrieval module, a report generation module, and a deep question answering module; 2) Use the content extraction module to extract core review dimensions from project application materials; 3) Based on the core review dimensions extracted in step 2), use the feature mining module to mine the technical keyword sequence in the project application materials; 4) The technical keyword sequence is vectorized using the literature retrieval module, and combined with the similarity algorithm, several technical comparison documents are retrieved from the technical literature database; 5) Based on the project application materials and the technical comparison literature obtained in step 4), use the report generation module to generate an AI review and evaluation report for the project application materials.

2. The method for assisting in project approval according to claim 1, characterized in that, The content extraction module is used to accurately locate and extract core review dimension information from project application materials through a dual-track mechanism that combines regular expression rule matching with semantic recognition of large language models. The feature mining module is used to extract highly discriminative technical keyword sequences from core review dimension information by driving a large language model through prompt word engineering. The document retrieval module is used to vectorize the sequence of technical keywords into project technical feature vectors, and to use a vector similarity algorithm to calculate the similarity between the project technical feature vectors and various technical documents in the technical document database. After multi-dimensional weighted sorting, several technical comparison documents are selected. The report generation module is used to generate an in-depth review and evaluation draft of the project application materials based on technical comparison literature and project application materials through search-enhanced generation technology, and then format and package it into a standardized AI review and evaluation report. The deep question-and-answer module is used to respond to users' questions about the review report, and output professional answers with evidence citations by combining the associated knowledge base, so as to realize closed-loop verification and traceability of the review logic; The data output terminal of the content extraction module is connected to the data input terminal of the feature mining module, the data output terminal of the feature mining module is connected to the data input terminal of the document retrieval module, the data output terminal of the document retrieval module is connected to the data input terminal of the report generation module, and the data output terminal of the report generation module is connected to the data input terminal of the deep question answering module.

3. The method for assisting in project approval according to claim 1, characterized in that, In step 2), the core review dimensions are extracted in the following manner: 2-1) Based on the directory structure of the project application materials, use regular expressions to locate the target title of the project application materials, extract the text content of the target title, and determine whether the text content is complete. If it is incomplete, proceed to step 2-2). 2-2) Construct content extraction instructions, and input the project application materials and content extraction instructions into the large language model, and use the large language model to extract key paragraphs from the project application materials; 2-3) The text content extracted in step 2-1) is deduplicated and merged with the key paragraphs identified by the large language model to obtain the core review dimension content.

4. The method for assisting in project approval according to claim 1, characterized in that, In step 3), the sequence of technical keywords is obtained in the following manner: 3-1) Based on the core review dimensions extracted in step 2), use the prompt word engineering to construct technical feature extraction instructions, and drive the large language model as the technical feature extraction model; 3-2) Input the preset technical dimension guidance instructions into the technical feature extraction model, and use the technical feature extraction model to perform deep semantic analysis on the text content of the project application materials to obtain the technical keyword sequence of the project application materials.

5. The method for assisting in project approval according to claim 1, characterized in that, In step 4), the comparative technical documents are obtained in the following manner: 4-1) Map the sequence of technical keywords obtained in step 3) to a high-dimensional vector space to construct the project feature vector; 4-2) Using the preset patent knowledge base and the user-uploaded associated document knowledge base as the technical document library, the vector similarity algorithm is used to calculate the semantic correlation between the project feature vector and each technical document in the technical document library; 4-3) Based on semantic relevance, select multiple technical documents as review reference sources from high to low; 4-4) Based on technical relevance, publication time, and legal status, the technical documents in the review reference sources are sorted by weight for a second time to select a number of technical comparison documents.

6. The method for assisting in project approval according to claim 1, characterized in that, In step 5), the AI ​​review and evaluation report is generated in the following manner: 5-1) The selected technical comparison literature and project application materials are used as context and injected into the prompts of the large language model. The large language model is then guided to generate an in-depth review and evaluation draft of the project application materials through search enhancement generation technology. 5-2) Formatting constraint instructions are constructed based on the JSON Schema formatting protocol. The formatting constraint instructions are used to drive the large language model to transform the initial draft of the in-depth review and assessment into structured data that conforms to the preset field definitions. 5-3) Using a document rendering engine, structured data is mapped to report templates, and AI audit and assessment reports in DOCX format are automatically packaged and generated.

7. The auxiliary project approval method according to claim 6, characterized in that, The formatting constraint instructions include: Pre-defined structured data description document conforming to the JSON Schema protocol; The description document is incorporated into the prompt word instructions, constraining the large language model to output JSON format data that conforms to the field definition; The JSON formatted data is populated into a preset document template using a document rendering engine.

8. The method for assisting in project approval according to claim 1, characterized in that, It also involves a process of responding to user questions using a deep question-answering module, specifically including: (1) Perform semantic parsing on the user's input questions to extract the core question points; (2) Based on the core questions, re-search the text fragments in the technical comparison literature that are most relevant to the project application materials; (3) Input the user questions, project application materials and text fragments obtained in step (2) into the large language model, use the large language model to output professional answers, and mark the source information of the cited literature in the reply.

Images