Automatic bidding document writing method, system and equipment based on multi-agent collaboration

Abstract

The invention discloses a bidding document automatic writing method, system and device based on multi-agent collaboration, and relates to the field of natural language processing, and the method comprises the steps: employing a knowledge base agent group to carry out tree graph structured processing on an enterprise business support material; performing layout analysis on the bid invitation file and the project information by adopting a bid invitation analysis agent group, and constructing a scoring point list and a scoring point query statement; performing outline generation and chapter writing type judgment by adopting an outline agent group; performing evidence retrieval by adopting a retrieval agent group; performing content writing by adopting a writing agent group; a compliance agent group is adopted to carry out compliance verification and backflow; generating a bidding document text and paragraph-level evidence source information; wherein each agent group is obtained by performing role constraint and output format constraint construction through different cue words based on a large language model. The bidding document writing efficiency is improved, and personalized requirements of different projects can be met.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing, in particular to a bid automatic writing method, system and device based on multi-agent collaboration. BACKGROUND

[0002] The preparation of bidding documents needs to meet the following requirements: bidder qualification and compliance, evaluation and scoring, invalid and bid cancellation conditions, mandatory documents, clause risks, and fairness. The automatic generation technology cannot correspond to the specific scoring points and rejection items of the bidding documents closely. At the same time, the newly written content lacks clear sources and scoring item constraints, and is prone to form illusory content. Although a large amount of historical bid documents and qualification / performance / specification data are accumulated in the enterprise, the common processing method is still limited to full-text search or simple semantic retrieval, and a controllable structured knowledge base has not been formed. There are the following problems.

[0003] 1. Mismatch between structure and scoring. The general bid template and material library are difficult to correspond to the scoring details, rejection items, and process nodes of different projects one by one, resulting in insufficient coverage of important scoring items and clauses or misplacement of content.

[0004] 2. Unclear source and "illusion" risk. Automatic writing technology is popular, but the newly written content lacks evidence support and often lacks source evidence, making it difficult to be accepted by the evaluators; old content is difficult to perfectly reuse; and there is no targeted processing for the content that needs to be reorganized.

[0005] 3. Single use of knowledge base. The technology of recalling knowledge base fragments only by vector retrieval easily ignores directory / section level and metadata constraints (security level, region / industry, project type, etc.), and the retrieval results are discrete and have poor usability.

[0006] 4. Lack of bid list management. Mandatory documents such as bid letters, itemized bid tables, legal authorization, business deviation tables, personnel and qualifications are not formed into a traceable list, and are easily missed.

[0007] 5. Insufficient compliance check. There is a lack of verification links and repair reminder mechanisms for "invalid / bid cancellation conditions, forbidden and limited language, format hard requirements, clause risks, and fairness review".

[0008] 6. Work flow black box cannot be split. The generated content has poor explainability, and there is no reasonable splitting into bid writing agents and workflow strategies, making it difficult to monitor and timely correct. SUMMARY

[0009] The purpose of the present application is to provide a bid automatic writing method, system and device based on multi-agent collaboration, which can improve the writing efficiency of bid documents and adapt to the individual needs of different projects.

[0010] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for automatically writing tender documents based on multi-agent collaboration, including: Obtain supporting materials, tender documents, and project information for the enterprise's business operations; The enterprise's business support materials are processed using a knowledge base intelligent agent group to obtain an enterprise tender document knowledge base and a hybrid index. The tender analysis intelligent agent group is used to parse the layout of the tender documents and the project information, obtain multi-field parsing results, and construct a list of scoring points and the corresponding query statements for each scoring point; Based on the multi-field parsing results, the scoring point list, the query statements corresponding to each scoring point, the enterprise tender knowledge base, and the hybrid index, the outline intelligent body group is used to generate the outline and determine the chapter writing type to obtain a customized project outline and writing list. Based on the customized outline of the project, the writing checklist, the query statements corresponding to each scoring point, the corporate tender knowledge base, and the hybrid index, evidence retrieval is performed using a retrieval intelligence group to obtain chapter evidence; Based on the customized outline of the project, the writing checklist, the chapter evidence, and the scoring point list, a writing intelligence group is used to write the content and obtain chapter drafts. Based on the chapter draft, the chapter evidence, the scoring point list, the writing list, and the query statement set corresponding to each scoring point, a compliance intelligent agent group is used to perform compliance verification and feedback to obtain a compliant version. Based on the compliant version, the chapter evidence, and the customized project outline, generate the tender document text and paragraph-level evidence source information; Among them, the knowledge base intelligent agent group, the bidding analysis intelligent agent group, the outline intelligent agent group, the retrieval intelligent agent group, the writing intelligent agent group, and the compliance intelligent agent group are all based on a large language model and are constructed by using different prompt words to constrain roles and output formats.

[0011] Secondly, this application provides an automatic tender document writing system based on multi-agent collaboration, comprising: The data acquisition module is used to acquire enterprise business support materials, tender documents, and project information; The knowledge base construction module is used to perform tree-graph structured processing on the enterprise business support materials using knowledge base intelligent agents to obtain the enterprise tender document knowledge base and hybrid index. The bidding analysis module is used to parse the layout of the bidding documents and project information using a bidding analysis intelligent agent group, obtain multi-field parsing results, and construct a list of scoring points and the corresponding query statements for each scoring point; The outline generation module is used to generate an outline and determine the chapter writing type by using an outline intelligent body group based on the multi-field parsing results, the scoring point list, the query statement corresponding to each scoring point, the enterprise tender knowledge base and the hybrid index, so as to obtain a customized project outline and writing list. The evidence retrieval module is used to retrieve evidence based on the customized outline of the project, the writing checklist, the query statements corresponding to each scoring point, the enterprise tender knowledge base, and the hybrid index, using a retrieval intelligent agent group to obtain chapter evidence. The content writing module is used to write content using a writing intelligence group based on the customized outline of the project, the writing list, the chapter evidence, and the scoring point list, and to obtain chapter drafts. The compliance verification module is used to perform compliance verification and feedback based on the chapter draft, the chapter evidence, the scoring point list, the writing list and the query statement set corresponding to each scoring point, using a compliance intelligent agent group to obtain a compliant version. The tender document generation module is used to generate tender document text and paragraph-level evidence source information based on the compliant version, the chapter evidence, and the customized outline of the project. Among them, the knowledge base intelligent agent group, the bidding analysis intelligent agent group, the outline intelligent agent group, the retrieval intelligent agent group, the writing intelligent agent group, and the compliance intelligent agent group are all based on a large language model and are constructed by using different prompt words to constrain roles and output formats.

[0012] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for automatically writing tender documents based on multi-agent cooperation.

[0013] According to the specific embodiments provided in this application, this application achieves the following technical effects: By constructing a multi-agent collaborative system based on a large language model and constrained by prompt word roles and output formats, it realizes an intelligent, customized, and compliant upgrade of the entire tender document writing process. Specifically, relying on the tree-graph structured processing and hybrid index construction of the knowledge base agent group, it achieves efficient integration and accurate retrieval of enterprise business support materials; through the refined analysis of the bidding analysis agent group, it accurately extracts scoring points and query statements, ensuring that the tender document content is highly consistent with the bidding requirements. The multi-agent collaboratively operates according to the process of outline generation, evidence retrieval, content writing, and compliance verification. It not only ensures the rationality of the tender document structure through customized outlines and chapter writing planning, but also enhances the credibility of the content by tracing the source of evidence at the paragraph level. At the same time, the compliance verification and feedback mechanism effectively avoids non-compliance risks. The entire solution significantly reduces manual intervention, significantly improves the efficiency of tender document writing, and can adapt to the personalized needs of different projects. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating an automatic tender writing method based on multi-agent collaboration, provided as an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the functional modules of an automatic tender writing system based on multi-agent collaboration, provided as an embodiment of this application.

[0017] Figure 3 This is an architecture diagram of an automatic tender writing system based on multi-agent collaboration, provided as an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To improve the relevance, traceability, and compliance of tender document preparation, this application establishes the following objectives.

[0020] 1. Standardized Knowledge Base Construction: The enterprise tender document knowledge base is constructed by a knowledge base intelligent agent group, using a tree-graph structure: the tree carries the directory and chapter hierarchy (business / technical trunk and sub-chapter); the graph carries evidence, qualifications, performance, terms and relationships (validity period, scope of application, conflicts / dependencies, etc.) across documents / chaps.

[0021] 2. Multi-agent collaboration: The bidding analysis agent group analyzes the bidding documents, then the outline agent group generates a customized outline according to the bidding requirements and scoring points, and classifies each chapter into writing types (copy / reorganize / create); the knowledge base agent group and the retrieval agent group perform different graph / tree evidence-based retrieval according to the writing type; the writing agent group generates content; and the compliance agent group checks and verifies the generated content.

[0022] 3. Evidence Constraints and Source Tracing: The generated content can locate the source evidence in the knowledge base (document ID, page number / anchor point, issuance date / validity period, scope of application, confidentiality level, evidence type, version); prompts are also given when there is insufficient evidence in special cases.

[0023] 4. Output of Results and Sources: Generates the main text of the tender document, the correspondence of scoring items, and paragraph-level evidence source information, which facilitates review and further optimization of the tender document content.

[0024] This application is based on a "tree-structured enterprise tender document knowledge base". It focuses on key points such as basic information of tender documents, eligibility and compliance, review and scoring, invalid and rejected bids, required documents for bidding, tender document requirements, terms and risks, and fairness review. First, an outline is generated and the writing type of each chapter is determined. Then, a search is carried out according to the writing type. Evidence is traced throughout the process, and if compliance fails, it is checked and corrected.

[0025] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] In one exemplary embodiment, such as Figure 1 As shown, a method for automatically writing tender documents based on multi-agent collaboration is provided. This method is executed by computer devices, specifically by a terminal or server alone, or by both a terminal and a server. The method includes the following steps 101 to 108.

[0027] Step 101: Obtain the enterprise's business support materials, bidding documents, and project information. The project information includes the region, procurement method, and bid section type.

[0028] Supporting materials for a company's business include its historical tender documents, qualifications / licenses, performance / proof of business achievements, industry standards / norms, white papers, and commonly used templates.

[0029] Step 102: Use a knowledge base intelligent agent group to perform tree graph structuring on the enterprise business support materials to obtain the enterprise tender knowledge base and hybrid index.

[0030] In a specific application example, step 102 includes steps 21 and 22.

[0031] Step 21: The knowledge base intelligent body group sequentially performs page structure parsing and directory annotation, content block segmentation and metadata annotation on the enterprise business support materials, and constructs chapter trees with the business section and technical section as the main branches, and establishes a graph structure to obtain the enterprise tender document knowledge base.

[0032] The table of contents includes chapters, sub-chapter sections, numbering, and anchor points. Content blocks include paragraphs, tables, and attachments. Content block segmentation preserves hierarchy and referencing relationships. Metadata includes source, issuance date, validity period, applicable region, industry, confidentiality level, evidence type (qualification, performance, standard, plan, case), project type, and applicable conditions.

[0033] The technical section typically includes chapters on technical background, functional module design, implementation organization, schedule, quality assurance, security and contingency, Service Level Agreement (SLA) / operations, data security and privacy, interfaces and integration, and testing and acceptance as fixed entity types. The diagram structure establishes nodes and relationships for evidence, qualifications, performance, terms, and risks, and records version information.

[0034] The enterprise tender document knowledge base includes a chapter tree, a collection of content blocks with metadata, and a graph structure knowledge base.

[0035] The chapter tree records the chapter level, chapter number, chapter title, and mounting relationship between chapters and content blocks for commercial and technical bids.

[0036] The collection of content blocks with metadata includes paragraphs, tables, and appendix summaries extracted from various historical tenders, qualifications / certificates, performance materials, standards, etc. Each content block is associated with a unique semantic identifier and includes metadata such as source document identifier, chapter path, issuance date, validity period, applicable region / industry, project type, confidentiality level, and evidence type.

[0037] The graph-structured knowledge base abstracts documents, chapters, content blocks, evidence, qualifications, performance, terms, risk points, project types, etc., into entity nodes, and abstracts relationships into edges, forming a knowledge graph that can be persisted and queried in a graph database.

[0038] Step 22: Construct a hybrid index based on the enterprise tender document knowledge base. The hybrid index is used to locate and trace back knowledge within the enterprise tender document knowledge base. The hybrid index includes a semantic index, a tag index, and a graph index.

[0039] (1) Semantic Indexing: Each content block (including paragraphs, table units, and appendix summaries) and its corresponding chapter title are transformed using an open-source semantic representation vector model (such as bge-large-zh) to obtain text semantic vectors, and a unique semantic identifier is assigned to each content block; the "semantic vector-semantic identifier" pairs are stored in an open-source vector database (such as Milvus) to support nearest neighbor retrieval based on vector similarity. The semantic index is stored in the vector database, and after the query statement or rating point description is mapped to a vector, semantically similar semantic identifiers can be retrieved quickly.

[0040] (2) Tag Index: The tag index is stored in a key-value / inverted structure, with key-value pairs established using the metadata of the content block and evidence node (such as source type, evidence type, applicable region, industry, project type, issuance date, validity period, confidentiality level, etc.) as keys. Each tag value corresponds to a set of semantic identifiers, which are used to perform precise or range filtering on the semantic recall results.

[0041] (3) Graph Index: For entity nodes (projects, qualifications, performance, terms, chapters, etc.) and their relationships in the knowledge graph, the connection information between nodes is stored in the graph database (such as neo4j) in the form of adjacency list and relationship table, supporting multi-hop path query and pattern matching based on graph structure.

[0042] By combining semantic indexing, tag indexing, and graph indexing, candidate content blocks can be recalled first using semantic similarity during subsequent retrieval, and then fine-tuned through tag filtering and graph path constraints, thereby improving the relevance and usability of retrieval results.

[0043] In the hybrid index, all types of index entries are mapped to entities and content blocks in the enterprise tender knowledge base using semantic tags or graph node identifiers. This allows the retrieval agent group and the writing agent group to efficiently locate and trace back to the specific content and knowledge relationships in the enterprise tender knowledge base simply by querying the hybrid index.

[0044] Step 103: Use the bidding analysis intelligent agent group to parse the layout of the bidding documents and the project information, obtain multi-field parsing results, and construct a list of scoring points and the corresponding query statements for each scoring point.

[0045] In a specific application example, step 103 includes steps 31 and 32.

[0046] Step 31: The bidding analysis intelligent agent group sequentially performs the following analysis on the bidding documents and project information: basic project information analysis, project qualification requirements analysis, project review requirements analysis, bid document requirements analysis, invalid and rejected bid items analysis, required bid documents analysis, bidding document review and risk analysis, and fairness review risk analysis, to obtain multi-field analysis results.

[0047] The project qualification requirements include bidder eligibility and compliance review. Project evaluation requirements include scoring criteria, bid opening / evaluation / award, handling of special circumstances, award requirements, and contract award and performance. Bid document requirements include language, currency, units of measurement, document composition, document format, price, and validity period. Invalid and rejected bids include situations where they are prohibited and situations leading to rejection / invalidation. Required bidding documents include supporting documents, a bid letter, a breakdown of price quotations, authorization from the legal representative, a commercial deviation table, marked items (such as asterisks or triangles), and personnel requirements. Bid document review and risks include risks related to the bid bond, payment terms, intellectual property, and project duration.

[0048] The bidding analysis AI group includes AI agents capable of parsing the above information: project basic information parsing AI agent, project qualification requirement parsing AI agent, project review requirement parsing AI agent, bid document requirement parsing AI agent, invalid and rejected bid item parsing AI agent, required bid documents parsing AI agent, bidding document review and risk analysis AI agent, and fairness review risk analysis AI agent.

[0049] Step 32: Generate a list of scoring points and corresponding query statements for each scoring point based on the multi-field parsing results. Specifically, multiple sets of executable knowledge graph query statements aligned with the knowledge base are formed based on the multi-field parsing results. There are multiple sets of query statements corresponding to each scoring point.

[0050] Step 104: Based on the multi-field parsing results, the scoring point list, the query statements corresponding to each scoring point, the enterprise tender knowledge base, and the hybrid index, the outline intelligent body group is used to generate the outline and determine the chapter writing type to obtain the customized project outline and writing list.

[0051] In a specific application example, the outline agent group includes an outline writing agent, a type judgment agent, and a rating judgment agent. Step 104 includes steps 41 to 43.

[0052] Step 41: The outline writing agent aligns the multi-field parsing results with the chapter tree in the enterprise tender knowledge base based on the query statements corresponding to each scoring point and the hybrid index, generating a customized project outline to ensure that each chapter corresponds to the specific requirements and scoring points of the tender document.

[0053] Step 42: The type-determining agent determines the writing type for each chapter of the customized outline for the project based on the type determination. The writing types include copying, reorganizing, and creating new content.

[0054] Step 43: The scoring agent, based on the scoring point list and the writing type of each chapter in the customized project outline, determines the scoring point coverage status for each chapter, forming a writing list. The coverage status includes covered, pending coverage, conflicting, and placeholder. The writing list includes the writing type and coverage status of each chapter in the customized project outline.

[0055] Step 105: Based on the customized project outline, the writing checklist, the query statements corresponding to each scoring point, the enterprise tender knowledge base, and the hybrid index, a retrieval agent group is used to perform evidence retrieval to obtain chapter evidence. Chapter evidence includes document number, page number / anchor point, issuance date / validity period, scope of application, confidentiality level, evidence type, conflict marker, and version.

[0056] In a specific application example, the retrieval agent group includes a copying retrieval agent, a reorganizing retrieval agent, and a newly created retrieval agent. Step 105 includes steps 51 to 53.

[0057] Step 51: For chapters in the customized project outline whose writing type is copy, based on the query statements corresponding to each scoring point and the hybrid index, retrieve the corresponding original text in the enterprise tender knowledge base to obtain the corresponding chapter evidence. This step is executed by a copy-type retrieval agent.

[0058] Specifically, for the content in the writing checklist, based on the query statements corresponding to each scoring point, the original text evidence that can be directly cited is retrieved first in the specific content nodes of the chapter; when there is multiple pieces of evidence on the same topic, the best one is selected according to weight and timeliness.

[0059] Step 52: For chapters in the customized project outline whose writing type is "reorganization," perform a multi-hop search on the graph structure in the enterprise tender knowledge base and aggregate the corresponding chapter evidence. This step is executed by a reorganization-type retrieval agent.

[0060] Specifically, multi-hop graph search is performed through a recombinant retrieval agent to aggregate evidence from multiple sources (solutions, indicators, cases, clauses) on the same writing topic. Theme aggregation and deduplication are performed, and when there is a source conflict, a "to be verified" flag is set.

[0061] Step 53: For chapters in the customized project outline whose writing type is "New", create corresponding chapter evidence based on standard specification clauses, general solution templates, and similar project paths as reference arguments. This step is performed by a new-type retrieval agent.

[0062] Specifically, the newly developed search agent retrieves standard / normative clauses, general solution templates, and similar project paths as reference arguments, and pre-sets "no direct evidence" and placeholder strategies.

[0063] Step 106: Based on the customized project outline, the writing checklist, the chapter evidence, and the scoring point list, a writing agent group is used to write the content, resulting in a chapter draft. The chapter draft includes paragraph-level writing content and corresponding sources of evidence.

[0064] In a specific application example, the writing agent group includes a copying writing agent, a reorganizing writing agent, and a new writing agent. Step 106 includes steps 61 to 63.

[0065] Step 61: For chapters in the customized project outline whose writing type is copy, directly quote the original text from the chapter evidence to obtain the corresponding chapter draft. This step is executed by a copy-type writing agent. Specifically, the copy-type writing agent directly quotes compliant original text from the chapter evidence, allowing for formatting and minor adjustments to the expression, while recording the source of the paragraph / chapter evidence.

[0066] Step 62: For chapters in the customized project outline whose writing type is "reorganization," the multi-source evidence is reorganized and polished in a targeted manner based on the chapter evidence and the scoring point list to obtain the corresponding chapter draft. This step is executed by a reorganization-type writing agent. Specifically, the reorganization-type writing agent reorganizes and polishes the multi-source evidence in a targeted manner based on the chapter evidence and the scoring point list, while recording the source of the paragraph / chapter evidence.

[0067] Step 63: For chapters in the customized outline of the project that are classified as "New Writing," draft the corresponding chapter based on the reference arguments. This step is performed by a New Writing agent. Specifically, when there is no direct evidence and new writing is permitted, the New Writing agent drafts based on reference arguments, clearly stating the reference basis and assumptions; if the writing information is insufficient, it prompts that supplementary content or adjustments to the outline are needed.

[0068] Step 107: Based on the chapter draft, the chapter evidence, the scoring point list, the writing list, and the query statement set corresponding to each scoring point, a compliance intelligent agent group is used to perform compliance verification and backflow to obtain a compliant version.

[0069] In a specific application example, the compliance agent group includes a content inspection agent, a compliance inspection agent, and an evidence inspection agent. Step 107 includes steps 71 to 74.

[0070] Step 71: Based on the list of scoring points, the query statements corresponding to each scoring point, and the writing list, perform a content check on the chapter draft and generate a content check result. This step is executed by a content checking agent, which checks the chapter content of the query statements corresponding to each scoring point and the writing list.

[0071] Step 72: Perform a compliance check on the chapter draft based on the query statements corresponding to each scoring point and the writing checklist, and generate a compliance check result. This step is performed by a compliance check agent. The compliance check agent checks the compliance items of the query statements and writing checklist corresponding to each scoring point, including rejection trigger conditions, eligibility, scoring matching, required bidding documents, prohibited and restricted terms and formats, and review of clause risks and fairness.

[0072] Step 73: Based on the evidence in the chapter, check whether the evidence metadata in the chapter draft is allowed to be used, whether it is within its validity period, and whether it is within its scope of use, to obtain the evidence check result. This step is performed by an evidence checking agent, which checks whether the evidence metadata is allowed to be used, whether it is within its validity period, and whether it is within its scope of application.

[0073] Step 74: Based on the content inspection results, the compliance inspection results, and the evidence inspection results, generate an issue list and generate remediation suggestions based on the issue list. Adjust the project customization outline, the writing list, the chapter evidence, and / or the chapter draft, i.e., return to the iterations of steps 104 to 106 to obtain a compliant version.

[0074] Step 108: Generate the tender document text and paragraph-level evidence source information based on the compliant version, the chapter evidence, and the customized project outline.

[0075] Specifically, by merging and formatting compliant versions, chapter evidence, and customized project outlines, the tender document text and paragraph-level evidence source information are output. Paragraph-level evidence source information includes document number, page number / anchor point, issuance date / validity period, scope of application, evidence type, version, and writing style.

[0076] The knowledge base intelligent agent group, the bidding analysis intelligent agent group, the outline intelligent agent group, the retrieval intelligent agent group, the writing intelligent agent group, and the compliance intelligent agent group are all based on a large language model and are constructed by using different prompt words to constrain roles and output formats.

[0077] In this application, each agent is based on the same or a set of large language models, and role constraints and output format constraints are implemented through the Prompt project. Each agent includes system prompts, task prompts, and output format constraints. The system prompts define the agent's role, responsibilities, and overall behavioral guidelines. The task prompts, combined with the current project context (such as project type, industry, or tender details), generate specific tasks. The output format constraints require the model to strictly output JSON or a specific structure for easy parsing by subsequent modules.

[0078] Based on the same inventive concept, this application also provides a system for implementing the methods described above. The solution provided by this system is similar to the solution described in the methods above; therefore, specific limitations in one or more system embodiments provided below can be found in the limitations of the methods described above, and will not be repeated here.

[0079] In one exemplary embodiment, such as Figure 2 As shown, an automatic tender document writing system based on multi-agent collaboration is provided, including: a data acquisition module 201, a knowledge base construction module 202, a tender analysis module 203, an outline generation module 204, an evidence retrieval module 205, a content writing module 206, a compliance verification module 207, and a tender document generation module 208.

[0080] The data acquisition module 201 is used to acquire enterprise business support materials, tender documents and project information.

[0081] The knowledge base construction module 202 is used to perform tree-graph structured processing on the enterprise business support materials using knowledge base intelligent agents to obtain the enterprise tender knowledge base and hybrid index.

[0082] In a specific application example, the knowledge base agent group includes the following agents: The document preprocessing and layout parsing intelligent agent is used to extract text, paginate, and recognize the layout structure (titles, body text, tables, headers and footers, etc.) of PDF / Word / scanned documents, and output a list of structured document blocks.

[0083] The content block segmentation agent is used to segment a list of structured document blocks into paragraph-level or table-level content blocks based on line spacing, indentation, punctuation, semantic coherence, etc., and output a list of content blocks.

[0084] The metadata annotation agent is used to identify tags such as source type, evidence type, validity period, applicable region / industry, project type, and confidentiality level for each content block, and output the content block with metadata.

[0085] The chapter tree building agent is used to identify the table of contents, chapter titles and levels, build the chapter tree, mount the content blocks with metadata to the corresponding chapter nodes, and output the chapter tree structure.

[0086] Knowledge graph construction agents are used to extract entities such as documents, chapters, evidence, qualifications, performance, and clauses as graph nodes, identify the relationships between them, and output a graph-structured knowledge base.

[0087] An index-building agent is used to construct semantic vector indexes, tag inverted indexes, and graph indexes for content blocks with metadata and nodes in a graph-structured knowledge base.

[0088] The aforementioned agents are linked together in the order of "document preprocessing → content block segmentation → metadata annotation → chapter tree construction → graph construction → index construction", and are controlled and executed by the orchestration and scheduling agent.

[0089] The bidding analysis module 203 is used to parse the bidding documents and project information using a bidding analysis intelligent agent group, obtain multi-field parsing results, and construct a list of scoring points and query statements corresponding to each scoring point.

[0090] In a specific application example, the bidding analysis agent group includes the following agents: The project basic information parsing intelligent agent is used to extract basic information such as project name, purchaser, budget amount, implementation period, and location, and output the project basic information.

[0091] The project qualification requirement parsing agent is used to identify qualification clauses such as bidder qualification conditions, personnel qualifications, and financial requirements, and output the project qualification requirements.

[0092] The project review requires a parsing agent to parse the scoring sheet, extract the scoring item names, scores, scoring rules, supporting material requirements, etc., and output a list of scoring points.

[0093] The tender document requirements parsing intelligent agent is used to identify the composition, chapter requirements, file format, language, currency, validity period, and other contents of the tender document, and output the tender document requirements.

[0094] The invalid and rejected bid parsing agent is used to identify clauses such as "considered invalid" and "will be rejected" from the text and output invalid and rejected bids.

[0095] The required documents parsing agent is used to parse the letters, forms, and supporting documents that must be submitted, and output the required documents for bidding.

[0096] The tender document review and risk analysis agent and the fairness review risk analysis agent are used to mark high-risk or suspected unfair terms related to payment terms, construction period, liability for breach of contract, intellectual property, etc.

[0097] The outline generation module 204 is used to generate an outline and determine the chapter writing type based on the multi-field parsing results, the scoring point list, the query statement corresponding to each scoring point, the enterprise tender knowledge base, and the hybrid index, using an outline intelligent body group to obtain a customized project outline and writing list.

[0098] In a specific application example, the outline agent group includes the following agents: The outline generation agent is used to generate a structured tender outline based on the scoring points, document requirements, and the chapter tree in the knowledge base, and output the chapter tree structure.

[0099] The chapter writing type determination agent is used to determine whether each chapter in the outline is "copy-type, reorganized-type, or newly created-type", and outputs a writing type table.

[0100] The rating point coverage mapping agent is used to map rating points to chapter nodes in a chapter tree structure, mark the coverage status of each rating point, and output a coverage relationship table.

[0101] The evidence retrieval module 205 is used to retrieve evidence based on the customized outline of the project, the writing checklist, the query statements corresponding to each scoring point, the enterprise tender knowledge base, and the hybrid index, using a retrieval intelligent agent group to obtain chapter evidence.

[0102] In a specific application example, the retrieval agent group includes the following agents: The copy-based retrieval agent is used to retrieve mature content paragraphs or templates that can be directly referenced from the knowledge base for copied chapters based on semantic indexing and tag indexing, and output evidence of copied chapters.

[0103] The recombinant retrieval agent performs multi-hop searches using graph indexes for recombinant sections, aggregating relevant evidence from multiple items and documents, and outputting evidence for the recombinant section.

[0104] The newly created search agent is used to retrieve reference evidence such as standards, specifications, and industry best practices for newly created chapters; if no direct evidence is available, it is marked as "insufficient evidence" and the reference evidence for the newly created chapters is output.

[0105] The content writing module 206 is used to write content using a writing intelligence group based on the customized outline of the project, the writing list, the chapter evidence, and the scoring point list, and to obtain chapter drafts.

[0106] In a specific application example, the writing agent group includes the following agents: A copy-writing AI agent is used to directly quote evidence text in copied chapters, perform slight formatting, and output draft fragments containing paragraph text and evidence citations.

[0107] The reorganization writing agent is used to merge, simplify, rearrange, and polish multiple evidence fragments in conjunction with scoring requirements for reorganized chapters, and output a draft of the reorganized chapter.

[0108] The new writing agent is used to generate innovative descriptions that meet the bidding requirements for new chapters, based on reference standards / specifications. If the evidence is insufficient, it will explicitly output a prompt, output a draft of the new chapter, and mark the corresponding paragraph as "no direct project-level evidence".

[0109] The compliance verification module 207 is used to perform compliance verification and feedback based on the chapter draft, the chapter evidence, the scoring point list, the writing list and the query statement set corresponding to each scoring point, using a compliance intelligent agent group to obtain a compliant version.

[0110] In a specific application example, the compliance agent group includes the following agents: The content coverage checking agent is used to check whether the chapter draft covers all the scoring points required in the relationship table and outputs a coverage check report.

[0111] The terms and format compliance check agent is used to check the format and compliance of drafts based on invalid and rejected bid items, required bid documents, bid document requirements, and a pre-set prohibited and restricted language rule library, and output a list of compliance issues.

[0112] The evidence validity checking agent is used to check whether the evidence cited in the draft is still valid and applicable to the current project, and outputs an evidence checking report, in accordance with the validity period and scope of application of the evidence.

[0113] The tender document generation module 208 is used to generate tender document text and paragraph-level evidence source information based on the compliant version, the chapter evidence, and the customized outline of the project.

[0114] In a specific application example, the orchestration and scheduling agent group runs throughout the system and mainly includes the following agents: The process orchestration agent is used to determine the calling order and concurrency strategy of each agent based on a predefined flowchart or configuration script.

[0115] The task tracking and retrying agent is used to track the execution status of each project instance, retry or rollback at faulty nodes, and ensure the stability of the overall process.

[0116] Log and audit agents are used to record summaries of each agent's inputs and outputs, as well as key decision paths, to facilitate post-event auditing and quality tracking.

[0117] To achieve decoupling and collaboration among multiple agents, this application abstracts cross-module data into several standardized structures, using JSON as the interface format. The orchestration and scheduling agent performs routing and state management based on task number and loading type.

[0118] Among them, the knowledge base intelligent agent group, the bidding analysis intelligent agent group, the outline intelligent agent group, the retrieval intelligent agent group, the writing intelligent agent group, and the compliance intelligent agent group are all based on a large language model and are constructed by using different prompt words to constrain roles and output formats.

[0119] like Figure 3As shown, the automatic tender writing system based on multi-agent collaboration provided in this application adopts a multi-agent collaborative architecture, which consists of the following layers.

[0120] The user interaction layer provides interfaces for configuring bidding projects, uploading bidding documents, uploading company information, monitoring the generation progress, and downloading results.

[0121] The agent orchestration and scheduling layer, implemented by the orchestration and scheduling agent group, manages the calling order, concurrency, state and exceptions of each agent in a unified manner, and is the "central hub" of multi-agent collaboration.

[0122] The business function intelligent agent layer contains multiple intelligent agent groups, and each intelligent agent group contains several specific intelligent agents. Each of them implements different roles on the same large model through the Prompt project: knowledge base intelligent agent group, bidding analysis intelligent agent group, outline intelligent agent group, retrieval intelligent agent group, writing intelligent agent group, and compliance intelligent agent group.

[0123] The model and algorithm support layer provides basic large model services, vector encoding models, relation extraction models, classification models and other underlying algorithm interfaces.

[0124] The data and indexing layer includes an enterprise tender knowledge base, hybrid indexes (semantic indexes, tag indexes, graph indexes), intermediate result storage, and log audit storage, and provides a standardized interface protocol (JSON / table structure).

[0125] The following example, using the "Comprehensive Security System Project for a Smart Park in a Certain City" as a scenario, illustrates the complete processing flow and intermediate result form of the automatic bid writing method based on multi-agent collaboration provided in this application in a real project.

[0126] (a) Project scenario and input instructions.

[0127] Project Background: A municipal park management committee issued a public tender notice for the "Smart Park Integrated Security System Construction Project". The main construction contents include: video surveillance subsystem, access control subsystem, perimeter protection, platform software, operation and maintenance services, etc., and the comprehensive scoring method is used for evaluation.

[0128] Available materials for the enterprise include the following tender documents and supporting materials accumulated internally: approximately 50 PDF / Word documents of commercial and technical bids for similar projects such as smart parks / safe cities / smart campuses in the past three years; scanned copies and electronic versions of the company's qualification certificates (security engineering enterprise qualification, ISO system certificate, ITSS operation and maintenance qualification, etc.); proof materials of typical project performance (scanned copies of bid-winning notices, contracts, acceptance reports, etc.); industry standards and specifications (such as GA / T, GB / T series security standards, etc.); and the company's internally unified technical solution template, implementation plan template, and service plan template.

[0129] Input data includes: Tender document: ZB_2025_SmartPark.pdf; Project information: Region is "a certain city", Procurement method is "open tender", Bid section type is "security engineering + software platform"; Collection of the company's historical tender documents and supporting materials.

[0130] (ii) Construction of a standardized knowledge base.

[0131] (1) Page layout analysis and chapter tree construction.

[0132] Batch parsing of enterprise historical tender documents and supporting materials: automatically identifying commercial / technical tenders and extracting table of contents, chapter numbers, titles and page number anchors.

[0133] Construct a unified chapter tree: the business section includes company overview, qualifications and performance, response to business terms, quotation explanation, etc.; the technical section includes project overview, understanding of requirements, overall architecture, video surveillance solution, access control and perimeter protection, system integration, implementation plan, operation and maintenance and service, data security, etc.

[0134] (2) Content block segmentation and metadata annotation.

[0135] The document is split into content blocks: for example, a certain paragraph of text in “3.2.1 Overall Design of Video Surveillance System” is assigned content_id=C_000123, where content_id represents the content number and is marked with the following information.

[0136] Source document: TechBid_2023_SmartCampus.pdf.

[0137] Chapter path: Technical specifications / Video surveillance solution / System overall design.

[0138] Evidence type: Scheme description.

[0139] Project type: Smart campus / industrial park.

[0140] Date of issuance: 2023-05-16, Valid until: 2026-05-15.

[0141] Applicable region: East China; Confidentiality level: Internal.

[0142] (3) The entity nodes of the graph structure include: ProjectType: Smart Park; Evidence:C_000123; SectionNode: Video Surveillance Solution - Overall Design; Standard: GA / T 367-202X Technical Requirements for Video Security Monitoring Systems.

[0143] The relational edges in a graph structure include: C_000123 maps_to_section video surveillance solution - overall design; C_000123 applies to ProjectType: Smart Park; C_000123 evidenced_by Standard:GA / T 367-202X; C_000123 valid_during 2023-05-16~2026-05-15.

[0144] Among them, ProjectType represents the project type, Evidence represents evidence, SectionNode represents the section node, Standard represents the standard, maps_to_section represents mapping to the section, applies_to ProjectType represents the applicable project category, evidenced_by Standard represents the evidence source standard, and valid_during represents the validity period.

[0145] (4) Create a hybrid index.

[0146] Semantic indexing: Encode all content blocks and chapter titles, write them into a vector database, and obtain a vector-semantic identifier mapping.

[0147] Tag Index: Create a key-value index, for example: key = Project Type: Smart Park → {C_000123, C_000456, ...}; key = Evidence Type: Performance Certificate → {E_000010, E_000015, ...}. Here, C_000123 and C_000456 represent two different content blocks, and E_000010 and E_000015 represent two different pieces of evidence.

[0148] Graph index: Stores the adjacency list of the graph structure in the graph database to support subsequent multi-hop queries.

[0149] The final output includes a corporate tender document knowledge base and a hybrid index, which can be used in subsequent steps.

[0150] (III) Analysis of the tender document.

[0151] (1) Page layout analysis and structured extraction. The ZB_2025_SmartPark.pdf was analyzed in a structured manner to obtain the basic information of the project, including the project name, funding source, bidder qualification requirements, contract period, budget amount, scoring method, etc.; chapter division: such as "Chapter 2 Instructions to Bidders", "Chapter 4 Evaluation Method", "Chapter 5 Contract Terms", "Technical Requirements", etc.

[0152] (2) Example of multi-agent analysis results.

[0153] The project qualification requirements analysis for intelligent agents identify that: "There must be at least one security project performance of smart park or safe city with an amount of not less than 8 million yuan in the past three years"; "It is necessary to have a Class A security engineering enterprise qualification, which is within the validity period", etc.

[0154] The project review requirements are analyzed using a "comprehensive scoring method": business aspect 30 points, technical aspect 60 points, and price 10 points; within the technical aspect, "performance of the video surveillance system and rationality of the solution" 20 points, "platform software functionality and visualization capabilities" 20 points, and "implementation organization and operation and maintenance services" 20 points.

[0155] The intelligent analysis of invalid and rejected bids revealed the following: "Bids that fail to provide a copy of the security engineering enterprise qualification certificate or whose certificate has expired will be rejected"; "Bids that fail to provide performance proof materials as required will be considered invalid."

[0156] (3) Example of scoring point list and knowledge graph query statement.

[0157] Example of a scoring point: The video surveillance system supports ≥2000 high-definition access channels and has intelligent analysis functions such as face recognition and behavior analysis. If this condition is met, a full score of 20 points will be awarded.

[0158] The corresponding knowledge graph query statement is as follows: Nodes: Section: Technical Requirements Document / Video Surveillance System; Concept: Video Access Capability; Concept: Intelligent Analysis Function (Facial Recognition / Behavioral Analysis). Here, Section represents a chapter, and Concept represents a concept.

[0159] Relationship: Section->Concept(requires); Concept->Evidence(evidenced_by). `requires` represents the requirement, `Evidence` represents the evidence, and `evidenced_by` represents the source of the evidence.

[0160] Query constraint: ProjectType=Smart Park.

[0161] The validity period includes the bidding period.

[0162] Evidence type ∈ {Solution, Parameter Indicator, Success Case}.

[0163] The final output includes a multi-field parsed structure, a list of scoring points, and the corresponding query statement.

[0164] (iv) Outline generation and chapter writing type determination.

[0165] (1) The outline writing agent aligns the query statement with the chapter tree to generate a customized outline for the project, for example: Chapter 1 Project Overview.

[0166] Chapter Two: Construction Goals and Needs Analysis

[0167] Chapter 3 Overall Architecture Design.

[0168] Chapter Four: Video Surveillance System Solution

[0169] Chapter 5 Access Control and Perimeter Protection Solutions.

[0170] Chapter Six: Platform Software and Data Sharing

[0171] Chapter 7 Implementation Organization and Schedule.

[0172] Chapter 8 Operation and Maintenance Services and SLAs.

[0173] Chapter Nine: Risk Analysis and Countermeasures.

[0174] Each chapter is linked to the corresponding rating points.

[0175] (2) The type judgment agent labels each section: "Company and project experience" → The company has a mature template, so it is judged as [copy type]; "Overall architecture of video surveillance system" → Historical projects can be referenced but need to be adjusted according to the needs of this project, so it is judged as [restructuring type]; "Future expansion planning and innovative application of the park" → New ideas based on the characteristics of this project, so it is judged as [new type].

[0176] Examples of some entries in the writing checklist are shown below.

[0177] ① Chapter 4.2 "Video Access Capabilities and Scalability": Coverage of scoring points: {C1, C2}; Writing type: Reorganization type; Coverage status: C1 = to be covered, C2 = partially covered.

[0178] ② Chapter 7.1 “Implementing the Organizational Structure”: Coverage of scoring points: {C10}; Writing type: Copying; Coverage status: C10 = Covered.

[0179] The final output includes a customized project outline and writing checklist.

[0180] (v) Evidence retrieval.

[0181] Example of copy-based retrieval: For the parts marked as "copy-based" in the writing checklist (such as standard service commitments, safety management systems, etc.), the copy-based retrieval agent, based on the constraints in the query statement (such as service level, response time), retrieves the original text that best matches the scoring requirements of this project from historical winning bids. It prioritizes selecting case content blocks from the most recent 2 years, projects with a similar scale to this project (amount ≥ 8 million), and similar applicable regions (a certain city or surrounding provinces) to form chapter evidence.

[0182] Example of Recombinant Retrieval: The recombinant retrieval agent performs a multi-hop search along the graph structure: starting from video access capabilities, it searches historical solutions for parameter tables regarding "maximum number of access channels," successful case studies showing actual "≥2000 access channels," and industry standards for access capabilities and bitrate control. Multiple pieces of evidence are aggregated: E1: Solution parameter table, stating "the system supports a maximum of 5000 HD access channels"; E2: A smart city project, with an acceptance report recording 4000 access channels; E3: Relevant national standard clauses. If different versions of system parameters exist (e.g., the old version has 3000 channels, the new version has 5000 channels), a conflict is marked, and a "to be verified" label is set. Here, E1, E2, and E3 represent three pieces of evidence.

[0183] Example of a newly created search: For the section "Future Expansion Planning and Innovative Applications of the Park," the newly created search agent primarily searches for: industry white papers on the development trends of smart parks, internal company research reports on innovative applications, and ideas from similar projects that are not directly included in the tender documents but are in the internal solution database. This evidence is marked as "Reference Basis" and included in the chapter evidence, while simultaneously marked with a placeholder "No direct bidding evidence."

[0184] The final output is a chapter of evidence, which records: evidence number, source document, page number or anchor point, issuance date / validity period, scope of application, evidence type, conflict marker and version information.

[0185] (vi) Content writing.

[0186] Taking the "Chapter 4 Video Surveillance System Solution" as an example, the writing agent generates chapter drafts under evidence constraints.

[0187] Example of copy writing: For the "Service Response Time Commitment" section, the copy writing agent directly adopts the wording that has been verified and complied in historical winning projects, and only makes parameterized replacements in fields such as project name and service time limit.

[0188] Example of restructured writing: For the section on "Video Access Capability and Scalability": Extract the parameter statement "supports up to 5000 HD access channels" from E1; extract the case study "4000 channels have been verified in a smart city project in a certain city" from E2; extract key standard clauses from E3 as compliance basis; the restructured writing agent integrates them into a logically clear description: first, state that the project plans to access ≥2000 channels and reserves the capability to expand to 5000 channels, and then support this with successful cases and standard clauses. Each paragraph records the corresponding E1 / E2 / E3 evidence citation list.

[0189] Example of a new writing approach: In the "Innovative Application Planning" section, the new writing agent refers to the industry trends and company research ideas in the chapter evidence to generate a description of AI-enabled scenarios tailored to the characteristics of this park (such as AI algorithms identifying the density of people and traffic flow in the park). At the same time, it states in the internal annotation: This paragraph is constructed based on references and reasonable assumptions, and there are no ready-made achievements or standards to directly support it. If necessary, it can trigger manual supplementation of materials or adjustment of the outline.

[0190] The final output is a chapter draft, with each section linked to the corresponding source of evidence.

[0191] (vii) Compliance verification and feedback.

[0192] Content inspection example: The content inspection agent compares the query statement with the writing checklist and checks: whether each rating point is clearly covered by the corresponding section in the chapter draft; whether there are any cases of "missing rating points" or "rating points contradicting the text description".

[0193] Compliance check examples: Check if all required bidding documents have been provided. Is the security engineering company's qualification certificate included in the draft appendix list? Do the performance verification materials match the requirement of "amount ≥ 8 million, within the last three years, project type is smart park / safe city". Check if any conditions for bid rejection have been triggered: If any required qualification is missing or expired, record the serious issue in the problem list and mark it as requiring a return to step (II) for supplementary certification. Check for prohibited and restricted terms and fairness issues: For example, whether expressions such as "the only one in the industry" or "the first in the country" appear, which may be considered unfair competition, and provide replacement suggestions.

[0194] Example of evidence inspection: Check whether the validity period of evidence such as E1 / E2 / E3 covers the bid deadline; if it is found that the issuance date of a certain performance certificate is too early and does not meet the "last three years" requirement, mark "evidence to be replaced" in the problem list and go back to step (V) to search again.

[0195] Example of iterative backtracking: For the issue "scoring points not covered" in the issue list, backtrack to step (iv) to adjust the outline and writing list. For the issue "evidence non-compliant", backtrack to step (v) to re-search for evidence; for the issue "description does not match scoring points", backtrack to step (vi) to adjust the writing content. After several rounds of iteration, a compliant version and issue list are obtained (if the remaining issues are only risk warnings).

[0196] (viii) Results and source output.

[0197] The following are some representative examples of agent prompts (excerpts): (1) The following is an example of the intelligent agent Prompt for analyzing the project review requirements in the tender analysis intelligent agent group.

[0198] The system prompt reads: "You are the project review requirement parsing agent. Your task is to identify and extract all review scoring items from the Chinese tender documents in a structured manner. The output must include: scoring item number, scoring item name, category, score, scoring rules, and required supporting materials. You are only responsible for understanding and structuring the data; you must not make any subjective evaluations or rewrite the data."

[0199] Example of task prompt: "[Input] is a passage of the original tender document, which may contain table text or paragraph descriptions:" Please extract all scoring items and output them as a JSON array as follows: [{"score_item_id":"Automatically generated unique number","name":"Scoring item name","category":"Business score / Technical score / Price score / Other","max_score":Score number,"rule":"Original scoring rule or a closely related summary","required_evidence":["Required supporting document 1","Required supporting document 2"]}]. If no scoring item is found, output []].

[0200] (2) The following is an example of the outline writing agent Prompt in the outline agent group.

[0201] Example system prompt: "You are an outline writing AI, skilled at designing tender document chapter structures based on scoring criteria and tender document requirements. The outlines you generate must: cover all scoring criteria; be compatible with common business / technical tender document structures; and facilitate a one-to-one correspondence between subsequent chapters and scoring criteria."

[0202] Example of a task prompt: "The rating items for the current project are as follows (JSON array):" The requirements for the sections / formats of the tender documents are as follows: Please output an outline JSON: {"nodes":[{"node_id":"Auto-generated","parent_id":null or parent node ID,"title":"Chapter Title","level":1 or 2 or 3,"write_type":"copy|recompose|new","score_item_ids":["List of associated score item IDs"]}]}". Where ScoreItems represents score items, DocRequirement represents the chapter / format requirements of the tender document, nodes represents nodes, node_id represents node number, parent_id represents parent node number, title represents chapter title, level represents level, write_type represents writing type, copy represents copy, recompose represents recombination, new represents new, and score_item_ids represents score item IDs.

[0203] (3) The following is an example of the recombinant search agent Prompt in the search agent group.

[0204] Example system prompt: "You are a reorganized retrieval agent. Given a chapter description and corresponding rating point information, you need to generate a retrieval intent suitable for querying in a graph-structured knowledge base. Your output will not return text directly, but will return retrieval instructions for subsequent retrieval modules to execute."

[0205] Example of a task prompt: "Chapter Information:" Corresponding scoring point information: Please generate a set of search intent JSON: {"semantic_queries":["Natural Language Semantic Search Statement 1","Statement 2"],"graph_queries":["Graph query mode 1, such as: find evidence related to smart city + operation and maintenance + emergency plan","..."],"label_filters":{"region":["{{Bidding Project Region}}"],"industry":["{{Industry}}"],"project_type":["{{Project Type}}"]}}". Where OutlineNode represents the output node, ScoreItemsForThisNode represents the scoring items for this node, semantic_queries represents the natural language semantic search statement, graph_queries represents the graph query, label_filters represents the label filter, region represents the region, industry represents the industry, and project_type represents the project type.

[0206] (4) The following is an example of the recombinant writing agent Prompt in the writing agent group.

[0207] Example system prompt: "You are a technical proposal writing AI (recombination type). Your responsibility is to merge, rearrange, and refine multiple pieces of historical evidence without fabricating facts, to create a proposal with a clear structure and closely aligned with the scoring criteria. Fabricating facts is prohibited. Moderate optimization of the wording is allowed, but the substantive meaning of the evidence must not be changed."

[0208] Example of a task prompt: "Chapter Information:" The following rating points need to be covered: The provided evidence set is as follows (JSON): Please generate a draft of the main text for this chapter, with the following requirements: respond to each scoring point; incorporate and integrate evidence as much as possible; use formal written Chinese; output by paragraph, and include a list of cited evidence_ids for each paragraph. Output format: [{"paragraph_id":"automatically generated","text":"paragraph text","evidence_refs":["EV-1001","EV-1002"]}]". Where EvidenceSet represents the set of evidence, evidence_id represents the evidence number, paragraph_id represents the paragraph number, text represents the text, and evidence_refs represents the evidence links.

[0209] (5) The following is an example of the content inspection agent Prompt in the compliance agent group.

[0210] Example system prompt: "You are the content coverage checking agent. Your task is to check whether the draft tender covers all scoring points and mandatory clauses. You need to mark the status as: covered, not covered, partially covered, etc., and provide the chapter-level location."

[0211] Example of a task prompt: "Draft content (organized by chapter):" Rating point coverage mapping: Please output the coverage check report: {"items":[{"score_item_id":"SC-001","status":"covered|partial|missing","related_outline_node_ids":["SEC-3-1"],"related_paragraph_ids":["P-0003","P-0004"],"remark":"Explanation of reasons or suggestions"}]}". Where DraftSections represents draft sections, CoverageMap represents the score point coverage map, items represents items, score_item_id represents the score item number, status represents the status, covered indicates covered, partial indicates partially covered, missing indicates not covered, related_outline_node_ids represents the associated output node numbers, related_paragraph_ids represents the associated paragraph numbers, and remark represents the remarks.

[0212] (6) The following is an example of the process orchestration agent Prompt in the orchestration and scheduling agent.

[0213] Example system prompt: "You are the orchestration and scheduling agent. You do not directly generate the tender document content, but are responsible for deciding which agent should be invoked next. You generate the next task list based on the current task status and the output of each module."

[0214] Example of a task prompt: "Current task status:" Completed modules: Issues to be addressed: Please output the list of tasks to be executed next: {"next_tasks":[{"to_agent":"OutlineGenerator|EvidenceRetriever|DraftWriter|ComplianceChecker","reason":"Trigger Reason","input_refs":["Related Object ID"],"priority":1}]}". Here, TaskState represents the task status, FinishedModules represents completed modules, ComplianceIssues represents pending issues, next_tasks represents the next task, to_agent represents the outline generation, OutlineGenerator represents the evidence retrieval, EvidenceRetriever represents evidence retrieval, DraftWriter represents collaboration, ComplianceChecker represents checking, reason represents the reason, input_refs represents input links, and priority represents the priority.

[0215] In summary, the beneficial effects of this application include at least the following points.

[0216] (1) The structure is highly relevant and the key scoring points are fully covered. The generated outline uses the "specific requirements and scoring points" extracted from the tender document as anchor points, and each item is mapped to a business or technical chapter. The writing is completed at the chapter level by category (copying, reorganizing, and creating). The compilation process always revolves around the scoring points, which can effectively reduce the omission and misalignment of scoring items.

[0217] (2) The content is reliable and the source is clear, avoiding "generating out of thin air". In the writing stage, factual content is written in accordance with evidence: the source evidence of the knowledge base is recorded for both copying and reorganization, and checks are carried out when the evidence is insufficient to avoid generating uncontrolled new content. In the end, each paragraph of the text can be traced back to a specific source, which is convenient for verification.

[0218] (3) Strong targeted restructuring capability, better reuse and consistency. While maintaining the stability of the chapter structure, historical cases, qualifications and key points of technical solutions are aggregated for the same topic, and the descriptions are restructured and polished according to the scoring orientation; similar projects can improve reuse efficiency and maintain the consistency of terminology and standards by following the restructuring writing strategy.

[0219] (4) Improved efficiency and reduced rework. First, define the outline and scoring to cover the writing checklist, then categorize and search for writing materials, significantly reducing large-scale revisions later. When evidence is insufficient, the intelligent agent provides timely analysis and reminders, triggering supplementary content prompts or outline revision prompts, avoiding concentrated rework close to submission. The overall cycle is more predictable, and the generation process can be monitored and adjusted at any time.

[0220] (5) Compliance risks are controllable and the process is observable. Continuous verification is carried out around key points such as qualification assessment, review and scoring, invalid and rejected bids, required bidding documents, prohibited and restricted terms and formats, and terms and fairness; problems directly point to the locations that need to be supplemented or modified. The grasp of rigid terms and procedural requirements is more stable, and the phased results are more transparent.

[0221] (6) Applicable scope and timeliness management are in place. The knowledge base uniformly records meta-information such as applicable region / industry, validity period, and confidentiality level at the evidence level, which can be used for filtering and inspection, avoiding the citation of expired, unauthorized, or inappropriate evidence.

[0222] (7) Low expansion and maintenance costs. For the existing enterprise tender document knowledge base, adding new information only requires completing the table of contents and metadata annotation to enter the subsequent process; the nodes of common technical chapters and business chapters can directly reuse the knowledge base entity types, which is also convenient for subsequent maintenance.

[0223] (8) Smooth cross-project reuse. For the existing enterprise tender document knowledge base, when facing similar industries or projects, the existing knowledge base and intelligent agent writing process can be quickly reused, shortening the preparation time while maintaining quality.

[0224] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0225] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0226] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0227] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0228] 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. When executed, the computer program 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).

[0229] 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, etc., and are not limited to these.

[0230] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0231] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for automatic tender writing based on multi-agent collaboration, characterized in that, The method comprises: acquiring enterprise business support materials, bidding documents and project information; adopting a knowledge base intelligent agent group to perform tree graph structured processing on the enterprise business support materials to obtain an enterprise bidding document knowledge base and a hybrid index; adopting a bidding analysis intelligent agent group to perform layout analysis on the bidding documents and the project information to obtain multi-field analysis results, and to construct a scoring point list and a query statement corresponding to each scoring point; adopting an outline intelligent agent group to perform outline generation and chapter writing type determination according to the multi-field analysis results, the scoring point list, the query statement corresponding to each scoring point, the enterprise bidding document knowledge base and the hybrid index, to obtain a project customized outline and a writing list; adopting a retrieval intelligent agent group to perform evidence retrieval according to the project customized outline, the writing list, the query statement corresponding to each scoring point, the enterprise bidding document knowledge base and the hybrid index, to obtain chapter evidence; adopting a writing intelligent agent group to perform content writing according to the project customized outline, the writing list, the chapter evidence and the scoring point list, to obtain a chapter draft; adopting a compliance intelligent agent group to perform compliance verification and backflow according to the chapter draft, the chapter evidence, the scoring point list, the writing list and a query statement set corresponding to each scoring point, to obtain a compliance version; generating a bidding document text and paragraph-level evidence source information according to the compliance version, the chapter evidence and the project customized outline; wherein the knowledge base intelligent agent group, the bidding analysis intelligent agent group, the outline intelligent agent group, the retrieval intelligent agent group, the writing intelligent agent group and the compliance intelligent agent group are all based on a large language model, and are constructed by different prompt words for role constraint and output format constraint. 2.The method of claim 1, wherein, adopting a knowledge base intelligent agent group to perform tree graph structured processing on the enterprise business support materials to obtain an enterprise bidding document knowledge base and a hybrid index, comprising: adopting a knowledge base intelligent agent group to perform layout structure analysis and directory annotation, content block segmentation and metadata annotation on the enterprise business support materials in sequence, and to construct a chapter tree with a business part and a technical part as the main trunk, while establishing a graph structure, to obtain an enterprise bidding document knowledge base; the enterprise bidding document knowledge base comprises a chapter tree, a content block set with metadata and a graph structure knowledge base; constructing a hybrid index according to the enterprise bidding document knowledge base; the hybrid index is used for locating and tracing back knowledge in the enterprise bidding document knowledge base; the hybrid index comprises a semantic index, a label index and a graph index. 3.The method of claim 1, wherein, adopting a bidding analysis intelligent agent group to perform layout analysis on the bidding documents and the project information to obtain multi-field analysis results, and to construct a scoring point list and a query statement corresponding to each scoring point, comprising: adopting a bidding analysis intelligent agent group to perform project basic information analysis, project qualification requirement analysis, project review requirement analysis, bidding document requirement analysis, invalid bidding and bid item analysis, necessary bidding document analysis, bidding document review and risk analysis and fairness review risk analysis on the bidding documents and the project information in sequence to obtain multi-field analysis results; According to the multi-field analysis result, the score point list, the query statement corresponding to each score point, the enterprise bid knowledge base and the hybrid index, a set of outline intelligent agents is used for outline generation and chapter writing type judgment to obtain a project customized outline and a writing list, including:

4. The method for automatic tender writing based on multi-agent collaboration according to claim 1, characterized in that, According to the query statement corresponding to each score point and the hybrid index, the multi-field analysis result is aligned with the chapter tree in the enterprise bid knowledge base by the outline writing intelligent agent to generate a project customized outline; The type judgment intelligent agent judges the writing type of each chapter in the project customized outline; The score judgment intelligent agent judges the score point coverage state of each chapter in the project customized outline according to the score point list and the writing type of each chapter in the project customized outline to form a writing list. The writing list includes the writing type and coverage state of each chapter in the project customized outline; the writing type includes copying, reorganization and new creation. According to the project customized outline, the writing list, the query statement corresponding to each score point, the enterprise bid knowledge base and the hybrid index, a set of retrieval intelligent agents is used for evidence retrieval to obtain chapter evidence, including:

5. The method for automatic tender writing based on multi-agent collaboration according to claim 1, characterized in that, For the chapter with the writing type of copying in the project customized outline, the corresponding original text is retrieved in the enterprise bid knowledge base based on the query statement corresponding to each score point and the hybrid index to obtain the corresponding chapter evidence; 6. The method for automatic tender writing based on multi-agent collaboration according to claim 5, characterized in that, For the chapter with the writing type of reorganization in the project customized outline, the graph structure in the enterprise bid knowledge base is searched and the corresponding chapter evidence is aggregated; For the chapter with the writing type of new creation in the project customized outline, the corresponding chapter evidence is newly created based on the standard specification provisions, the general scheme template and the similar project path as reference arguments. According to the project customized outline, the writing list, the chapter evidence and the score point list, a set of writing intelligent agents is used for content writing to obtain a chapter draft, including: For the chapter with the writing type of copying in the project customized outline, the original text in the chapter evidence is directly quoted to obtain the corresponding chapter draft; 7. The method according to claim 5, wherein the method further comprises: For the chapter with the writing type of reorganization in the project customized outline, the multi-source evidence is specifically reorganized and polished according to the chapter evidence and the score point list to obtain the corresponding chapter draft; For the chapter with the writing type of new creation in the project customized outline, the corresponding chapter draft is written based on the reference arguments. According to the chapter draft, the chapter evidence, the score point list, the writing list and the query statement set corresponding to each score point, a set of compliance intelligent agents is used for compliance verification and backflow to obtain a compliance version, including: According to the score point list, the query statement corresponding to each score point and the writing list, the content of the chapter draft is checked to generate a content check result; 8.The method of claim 1, wherein, ​ ​ According to the query statements corresponding to each score point and the writing list, the chapter draft is checked for compliance items to generate a compliance checking result; According to the chapter evidence, whether the evidence metadata in the chapter draft is allowed to be available, is within a valid period, and is within a usage range is checked to obtain an evidence checking result; According to the content checking result, the compliance checking result, and the evidence checking result, a problem list is generated, and a repair suggestion is generated according to the problem list, and the project customized outline, the writing list, the chapter evidence, and / or the chapter draft are adjusted to obtain a compliance version.

9. A tender automatic drafting system based on multi-agent collaboration, characterized in that, The system performs the tender automatic writing method based on multi-agent collaboration in any one of claims 1-8, and the system comprises: A data acquisition module is configured to acquire enterprise business support materials, a tender document, and project information. A knowledge base construction module is configured to perform tree graph structural processing on the enterprise business support materials by using a knowledge base agent group to obtain an enterprise tender knowledge base and a hybrid index. A tender analysis module is configured to perform layout analysis on the tender document and the project information by using a tender analysis agent group to obtain a multi-field analysis result and construct a score point list and a query statement corresponding to each score point. An outline generation module is configured to generate an outline and determine a chapter writing type by using an outline agent group according to the multi-field analysis result, the score point list, the query statement corresponding to each score point, the enterprise tender knowledge base, and the hybrid index to obtain a project customized outline and a writing list. An evidence retrieval module is configured to perform evidence retrieval by using a retrieval agent group according to the project customized outline, the writing list, the query statement corresponding to each score point, the enterprise tender knowledge base, and the hybrid index to obtain chapter evidence. A content writing module is configured to perform content writing by using a writing agent group according to the project customized outline, the writing list, the chapter evidence, and the score point list to obtain a chapter draft. A compliance verification module is configured to perform compliance verification and backflow by using a compliance agent group according to the chapter draft, the chapter evidence, the score point list, the writing list, and a query statement set corresponding to each score point to obtain a compliance version. A tender generation module is configured to generate a tender text and paragraph-level evidence source information according to the compliance version, the chapter evidence, and the project customized outline. The knowledge base agent group, the tender analysis agent group, the outline agent group, the retrieval agent group, the writing agent group, and the compliance agent group are all based on a large language model, and are constructed by using different prompt words for role constraint and output format constraint.

10. A computer device comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the tender automatic writing method based on multi-agent collaboration in any one of claims 1-8.

Images