An automatic tender document generation method based on a large model combined with a multi-path recall knowledge base

By constructing a bidding document tree and intermediate table results, and combining them with a multi-path recall knowledge base, a draft bid is generated and its consistency is verified. This solves the problem of insufficient consistency in bid generation in existing technologies and achieves the accuracy and traceability of bid content.

CN123070282APending Publication Date: 2026-11-10BEIJING HKRSOFT TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611229393.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-13
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing automatic tender document generation technology has difficulty ensuring consistency between the main text clauses and the table items when processing tender documents, and the generated content is not consistent with the tender requirements, company information and supporting materials, resulting in a lack of traceability and accuracy in the generated tender document content.

Method used

By constructing a bidding document tree and intermediate table results, a bidding requirement parsing set is generated. Combined with a multi-path recall knowledge base, clause category determination and relationship node search are performed to generate a draft bid. Then, the source consistency is verified through a natural language reasoning chain, and finally, the final bid is generated.

Benefits of technology

It improves the traceability and accuracy of automatically generated tender documents, ensures consistency in clause responses, knowledge sources, and supporting materials within the chapter structure, and reduces issues of unfounded generation and miscitation of materials.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070282A_ABST
    Figure CN123070282A_ABST
Patent Text Reader

Abstract

This invention discloses an automatic bid document generation method based on a large model combined with a multi-path recall knowledge base, relating to the field of artificial intelligence technology. The method includes: hierarchically decomposing the chapter content and table items of the bidding documents, constructing a bidding document tree and intermediate table results, and extracting bidding elements to generate a bidding requirement parsing set; using the bidding requirement parsing set to classify, recursively abstract, and label entity relationships in the enterprise's bidding materials, constructing a bidding knowledge tree index and a bidding knowledge graph index; and calling the bidding document tree, bidding knowledge tree index, and bidding knowledge graph index to determine clause categories, find relationship nodes, extract summaries, and perform multi-path recall matching to generate a candidate knowledge set of clauses. This invention improves the traceability and review adaptability of the automatically generated bid document results by verifying the consistency of the source of the initial draft and revising it through dialogue.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base. Background Technology

[0002] With the development of electronic bidding, enterprise knowledge management, and natural language processing technologies, bid preparation has gradually evolved from manual data compilation and the use of fixed templates to an auxiliary preparation method based on document parsing, information extraction, knowledge base retrieval, and text generation. Current bid document processing typically extracts text from the main body, table of contents, tables, and attachments of the bidding documents, then combines this with historical bids, qualification certificates, project cases, product information, and technical solutions to form an enterprise knowledge base. The content of each section of the bid document is then generated through keyword retrieval, semantic retrieval, or large language models. Against the backdrop of the continuous maturation of artificial intelligence applications, generative pre-trained language models, vector retrieval, knowledge graphs, and multimodal content recognition are also increasingly being used in bidding clause understanding, bid document matching, and bid text generation scenarios.

[0003] While existing automated tender document generation technologies can improve text preparation efficiency to some extent, most solutions still rely heavily on template filling, single-path retrieval, or basic semantic matching, failing to adequately utilize the constraints between chapter levels, table items, clause sources, and supporting documents within the tender document. This is particularly problematic when the tender document includes qualification review forms, scoring sheets, technical parameter tables, price lists, and service commitment tables. Existing technologies are prone to issues such as discrepancies between text clauses and table items, unclear correspondence between company information and tender requirements, and a lack of traceable sources for generated paragraphs. Furthermore, existing large-scale model generation methods often generate the text first and then manually proofread it, making it difficult to define chapter paths, the location of supporting document citations, and response criteria before generation. This results in tender documents that, while linguistically complete, lack consistency with the tender clauses, company information, and supporting documents. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an automatic tender document generation method based on a large model combined with a multi-path recall knowledge base, which is used to solve the problem of inconsistent matching between clause responses and knowledge sources in automatic tender document generation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base, comprising: The content of the chapters and tables in the bidding documents are broken down hierarchically to construct a bidding document tree and intermediate table results, and bidding elements are extracted to generate a set of bidding requirements analysis. The bidding requirements parsing set is used to classify, recursively abstract, and identify entity relationships in the enterprise bidding materials, and to construct a bidding knowledge tree index and a bidding knowledge graph index. The tender document tree, tender knowledge tree index, and tender knowledge graph index are used to determine clause categories, find relationship nodes, extract summaries, and perform multi-way recall matching to generate a set of candidate clause knowledge. Based on the candidate knowledge set of clauses, intermediate results in tables, and format constraints, the mind tree chapter path is selected, and a forward search enhancement is triggered before paragraph generation to generate the first draft of the tender document; By using a natural language reasoning chain, the initial draft of the tender document, the set of tender requirements parsing, and the set of candidate clauses knowledge are verified for consistency and revised through dialogue to generate the final draft of the tender document.

[0007] As a preferred embodiment of the automatic tender document generation method based on a large model and a multi-path recall knowledge base described in this invention, the specific steps for constructing the tender document tree and intermediate table results are as follows: The layout of the bidding documents, including the table of contents, clause numbers, main text paragraphs, and appendix descriptions, is analyzed. A generative pre-trained language model is used to determine the semantic category and generate a bidding document tree. The system identifies row and column positions, completes semantic information for rows and columns, assigns items to their correct positions, and connects content across cells in the qualification review form, scoring form, technical parameter form, quotation list, and service commitment form in the bidding documents, generating intermediate table results.

[0008] As a preferred embodiment of the automatic tender document generation method based on a large model combined with a multi-path recall knowledge base as described in this invention, the specific steps for generating the tender requirement parsing set are as follows: Based on the bidding document tree and intermediate table results, extract the chapter bidding item set and table bidding item set respectively, and arrange the chapter bidding item set and table bidding item set into a bidding element chain according to the chapter source, clause source and table source; Establish requirement entries based on the source location and item category in the bidding element chain, and generate a bidding requirement parsing set.

[0009] As a preferred embodiment of the automatic bid generation method based on a large model combined with a multi-path recall knowledge base as described in this invention, the specific steps for constructing the bid knowledge tree index are as follows: The content of enterprise bidding documents is analyzed, and the purpose of each chapter is divided into chapters using the bidding requirements analysis set, generating chapter-based knowledge items; By using a recursive abstraction processing mechanism, chapter-based knowledge items are hierarchically divided, so that the bottom-level nodes retain the original knowledge content, the middle-level nodes form a classification summary, and the top-level nodes form an overview of the enterprise's capabilities. The retrieval path is established based on the summary hierarchy among the bottom-level nodes, middle-level nodes, and top-level nodes, and a bidding knowledge tree index is generated.

[0010] As a preferred embodiment of the automatic bid generation method based on a large model combined with a multi-path recall knowledge base as described in this invention, the construction of the bid knowledge graph index includes the following specific steps: The bidding entities are extracted from the bidding requirements parsing set and the bidding knowledge tree index, and the bidding entities are merged with the same name, divided into hierarchical and source locations to generate an entity node set; Determine ownership, proof, response, and reference relationships among entity node sets, and bind the corresponding relationships to node paths in the bidding knowledge tree index to generate a bidding knowledge graph index.

[0011] As a preferred embodiment of the automatic tender document generation method based on a large model and a multi-path recall knowledge base described in this invention, the specific steps for generating the candidate knowledge set of terms are as follows: The current requirement item is selected as the current response clause from the tender requirement parsing set. The chapter node and table source position of the current response clause are located by calling the tender document tree, the semantic category of the current response clause is determined, and a clause category label is generated. The entity node set in the bidding knowledge graph index is retrieved by using clause category tags. Enterprise data nodes that have proof relationships, response relationships, and reference relationships with the current response clause are filtered to generate a clause relationship node chain. By tracing back the top-level, middle-level, and bottom-level nodes in the bidding knowledge tree index using the clause relationship node chain, the enterprise capability overview, category summary, and original knowledge content are extracted to generate a clause knowledge tree summary. Starting from the entity node corresponding to the current response clause, search for adjacent enterprise data nodes, supporting material nodes, and project case nodes along the proof relationship, response relationship, and citation relationship in the clause relationship node chain, and retain the relationship type and source location between adjacent nodes to generate the clause graph neighborhood; Path compression and source sorting are performed on the neighborhood of the clause graph to obtain the entity path and clause graph summary connected to the current response clause; Based on the keywords, semantic content, and data type of the current response terms, keyword recall, semantic recall, and multimodal recall are performed respectively to obtain text matching knowledge, semantic matching knowledge, and supporting material knowledge; The knowledge tree summary of clauses, the graph summary of clauses, the text matching knowledge, the semantic matching knowledge, and the supporting material knowledge are arranged according to the chapter positions in the bidding document tree to form a multi-path recall knowledge base; Select knowledge content from the multi-path recall knowledge base that matches the clause category label, table source location, and entity path to generate a clause candidate knowledge set.

[0012] As a preferred embodiment of the automatic tender document generation method based on a large model combined with a multi-path recall knowledge base as described in this invention, the format constraints include the chapter order, heading level, table items, response criteria, and supporting material citation positions determined by the tender requirement parsing set and intermediate table results.

[0013] As a preferred embodiment of the automatic tender document generation method based on a large model and a multi-path recall knowledge base described in this invention, the specific steps for selecting the chapter path of the mind tree are as follows: The candidate knowledge set of clauses is used to match the chapter nodes in the bidding document tree to establish the corresponding chapter candidate paths. The table intermediate results and format constraints are combined to verify the position of table items, the position of supporting materials and the response scope in the chapter candidate paths, and generate a set of optional chapter paths. The tender document generation path is generated by using a mind tree reasoning process to deduce the clause continuation relationship, material reference relationship and paragraph order of the optional chapter path set, and selecting the chapter path deduction result that is consistent with the tender document tree, the candidate knowledge set of clauses and format constraints.

[0014] As a preferred embodiment of the automatic tender document generation method based on a large model combined with a multi-path recall knowledge base as described in this invention, the step of triggering look-ahead retrieval enhancement before paragraph generation includes the following specific steps: The tender document generation path is used to determine the current chapter to be generated and the position of the next paragraph, forming the semantics of the paragraph writing, and identifying clause names, technical matters, qualification matters, case matters and supporting material names to generate paragraph retrieval intent; Based on the candidate knowledge set of paragraph retrieval intent callback clauses, the bidding knowledge tree index, and the bidding knowledge graph index, the source content corresponding to the next paragraph position is obtained, and a paragraph source knowledge fragment is generated. The semantics of paragraph writing, knowledge fragments from paragraph sources, format constraints, and the path to generate the tender document are arranged into paragraph generation prompts, and the chapter titles, response items, cited materials, and paragraph order are limited. When the source knowledge fragment of a paragraph is consistent with the semantics of the paragraph, a formal paragraph is generated. When the source knowledge fragment of a paragraph lacks corresponding content of the clause, the candidate knowledge set of the clause is called again to supplement the source content and correct the semantics of the paragraph, and a revised paragraph is generated. Arrange the formal and revised paragraphs according to the tender document generation path to generate the initial draft of the tender document.

[0015] As a preferred embodiment of the automatic tender document generation method based on a large model combined with a multi-path recall knowledge base as described in this invention, the specific steps for generating the final tender document are as follows: The initial draft of the tender document is divided into response sentences for the qualification response, technical response, commercial response, quotation explanation, project implementation and after-sales service content, forming a sequence of response sentences; The natural language reasoning chain is used to determine the consistency between the response sentence sequence and the bidding terms, source content, knowledge tree summary and graph summary, and generate source consistency verification results. When the source consistency verification result indicates that the response sentence lacks source support, the candidate knowledge set of the clauses is invoked to supplement the corresponding source content and generate a supplementary response sentence. When the source consistency verification result indicates that there is a content difference between the response sentence and the tender terms, the corresponding response sentence is rewritten by calling the tender requirements parsing set and the clause candidate knowledge set, and a revised response sentence is generated. The target chapters in the initial draft of the tender document are located using user dialogue content, and the content of the target chapters is revised by combining supplementary response sentences and revision response sentences to generate the final draft of the tender document.

[0016] The beneficial effects of this invention are as follows: Instead of simply generating bid text using a large model, this invention first uses the intermediate results of the bidding document tree and tables to uniformly parse the main text clauses, table items, attachment descriptions, and their source locations into a unified bidding requirement parsing set. This parsing set then drives the formation of a bidding knowledge tree index and a bidding knowledge graph index from the enterprise's bidding data, enabling the enterprise's data to be organized and accessed according to the relationship of "bidding requirements—enterprise response content—supporting materials—citation location." Compared to existing bid generation methods that rely on template filling, keyword retrieval, or single-vector recall, this invention prioritizes knowledge content with proof relationships, response relationships, citation relationships, and complete source fields through a multi-path recall knowledge base. Furthermore, before paragraph generation, it utilizes mind tree chapter path selection and forward-looking retrieval to enhance the determination of chapter order, response scope, table location, and supporting material citation location, reducing unfounded generation, incorrect citation of materials, and omissions in clauses from the source. Furthermore, by using natural language reasoning chains to verify the consistency of sources between the initial draft of the tender and the revised content of the dialogue, the final tender is made consistent in terms of clause response, knowledge sources, supporting materials and chapter structure, thereby improving the traceability, accuracy and review adaptability of the automatically generated tender. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base.

[0019] Figure 2 This diagram illustrates the hierarchical breakdown of the tender documents and the generation of the tender requirements parsing set.

[0020] Figure 3 A schematic diagram illustrating the construction of the bidding knowledge index and the generation of the candidate knowledge set for clauses.

[0021] Figure 4 A flowchart for generating the initial draft of the tender document, verifying its source, and revising it through dialogue. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base, including the following steps: S1. Decompose the chapter content and table items of the bidding documents hierarchically, construct the bidding document tree and intermediate table results, extract the bidding elements, and generate a set of bidding requirements analysis.

[0026] S1.1. Receive the tender documents and determine their source. The tender documents are derived from electronic files downloaded from the electronic bidding platform, attachments provided by the tenderer, and scanned paper documents. Electronic files include parsable text files and formatted files, while scanned paper documents undergo text recognition processing to obtain parsable text content. The tender documents retain the document name, page number, attachment name, clause number, and table page number, so that the subsequently generated tender document tree, intermediate table results, and tender requirement parsing set can all return to their original positions in the tender documents. This prevents bidding elements from deviating from the clause page numbers and table positions, and facilitates the subsequent reference of tender requirements in the main body of the tender document.

[0027] The layout of the tender documents is analyzed, including the table of contents, clause numbers, main text paragraphs, and appendix descriptions. The main text, table of contents, tables, and appendix areas are divided according to headers and footers, title font size, numbering level, paragraph indentation, line break positions, and appendix identifiers. Table of contents titles, clause numbers, main text paragraphs, and appendix descriptions are mapped to the same hierarchical sequence, generating a tender chapter structure set. When the table of contents title differs from the main text title, the chapter hierarchy is determined by the clause number and title position within the main text. When appendix descriptions lack independent titles, they are included in the main text clauses that reference them. This layout-analyzed tender chapter structure set preserves the hierarchical relationships between chapter titles, clause paragraphs, and appendix descriptions. Compared to searching clauses solely by keywords, this method reduces parsing errors caused by qualification clauses, technical clauses, and commercial clauses being mixed in the same paragraph in complex tender documents.

[0028] S1.2. A generative pre-trained language model is used to determine the semantic categories of the bidding chapter structure set and generate a bidding document tree. The generative pre-trained language model adopts a Transformer-based generative pre-trained language model. The generative pre-trained language model receives the chapter titles, clause paragraphs, appendix descriptions and adjacent level texts from the bidding chapter structure set, and outputs the semantic categories corresponding to the project overview, qualification requirements, technical requirements, commercial requirements, quotation requirements, delivery requirements and review rules. When a chapter contains multiple semantic categories, the main category is determined by the chapter title and the sub-category is determined by the clause paragraph. The sub-category is then attached to the corresponding main category node to form a bidding document tree with parent nodes, child nodes, clause text and source page number.

[0029] The qualification review form, scoring form, technical parameter form, quotation list and service commitment form in the bidding documents are identified by row and column positions. For parsable text tables, the table structure is determined according to table lines, cell boundaries, row and column headers and the position of merged cells. For scanned tables, the cell text is first obtained through text recognition, and then the row and column structure is determined according to the horizontal and vertical dividing lines and the text arrangement position. For tables spanning multiple pages, the consistency of the first row header of the table on the next page with the row and column headers of the table on the previous page is compared, and consecutive pages of tables are grouped into the same table item.

[0030] S1.3. The table structure records are semantically completed, items are placed, and cross-cell content is connected through the table reasoning chain algorithm to generate intermediate table results. First, the row title, column title and cell content are combined into table item statements. Then, the superior items in the merged cells are extended to the corresponding sub-rows and sub-columns. Next, blank cells, cross-page continuation cells and multi-line description cells are added to the semantics of adjacent rows and columns. Finally, the review items, scoring items, parameter items, quotation items and service items are respectively included in the intermediate table results.

[0031] When semantically completing blank cells, the header of the same row and the header of the same column of the blank cell are inherited first. The inherited headers and headers are then combined with the content of adjacent non-empty cells to form the table item statement. When connecting tables across pages, the following rules apply: consistent headers, consistent number of columns, and consistent table names. Consistent headers mean that the column headers of the subsequent page are the same as those of the previous page or belong to the same merged header. Consistent number of columns means that the number of valid columns in the subsequent page is the same as that in the previous page. Consistent table names mean that the names of the qualification review table, scoring table, technical parameter table, quotation list, or service commitment table are the same for consecutive pages. When all three rules are met, the content of the subsequent page table is connected to the item of the previous page table. If any rule is not met, the subsequent page table is treated as a new table item and repositioned.

[0032] For statements such as "same as above", "see attachment", or "comply as required by the tender document" in the table, connect the corresponding cells to the items above in the same column, the attachment descriptions, or the corresponding clauses in the tender document tree to avoid the lack of context when extracting tender elements in the table.

[0033] Specifically, when the cell content is "Same as above", the connection is prioritized to the nearest non-empty cell above this cell in the same table and column, inheriting the content, source of the clause, and row and column position of the previous non-empty cell; when the cell content is "See attachment", the connection is prioritized to the attachment description referenced by the same clause, retaining the attachment name, attachment number, and clause number referencing the attachment; when the cell content is "Execute as required by the tender document", the connection is prioritized to the main text clause in the tender document tree corresponding to the row title, column title, and table name of this cell; when multiple candidate sources exist simultaneously, the final connection source is determined in the following order: source in the same table first, source in the same chapter second, and attachment source last. Source in the same table means that the candidate source is in the same qualification review table, scoring table, technical parameter table, quotation list, or service commitment table as the current cell; source in the same chapter means that the candidate source corresponds to the same chapter node in the tender document tree as the current cell; attachment source means that the candidate source comes from the attachment description or attachment file.

[0034] Based on the tender document tree, extract the chapter-based bidding item sets and the table-based bidding item sets based on the intermediate table results. Extract qualification, performance, personnel, and certificate items from the qualification requirement node in the tender document tree; extract technical parameters, equipment configuration, and implementation requirements from the technical requirement node; extract service response, delivery arrangements, and pricing constraints from the commercial requirement node; and extract scoring items and response format items from the review rule node, thus forming the chapter-based bidding item sets. Extract table-based bidding items from the qualification review table, scoring table, technical parameter table, pricing list, and service commitment table in the intermediate table results, thus forming the table-based bidding item sets. The chapter-based bidding item sets reflect the requirements of the main text clauses, while the table-based bidding item sets reflect the requirements of the table items. Both types of bidding items jointly cover the textual and table requirements in the tender documents, avoiding the omission of scoring items, parameter items, and pricing items due to only parsing the main text.

[0035] S1.4. Arrange the chapter-based and table-based bidding item sets into a bidding element chain according to chapter source, clause source, and table source. When the same bidding item appears in both the main text clause and the table item, retain the requirements in the main text clause and the filling position in the table item. When multiple table items in the same chapter correspond to the same main text clause, arrange the multiple table items under the same clause source according to the table page number and row / column position. When the appendix description limits the main text clause, use the appendix source as a supplementary source to the bidding element chain.

[0036] Based on the source location and item category in the bidding element chain, requirement entries are created to generate a bidding requirement parsing set. Each requirement entry includes a requirement category, requirement content, chapter source, clause source, table source, and response location. The requirement category corresponds to qualification requirements, performance requirements, personnel requirements, equipment requirements, pricing requirements, service response requirements, and format requirements. When a bidding element chain has both text and table sources, the requirement entry retains the text source as the basis for the requirement and the table source as the basis for completion. When a bidding element chain only has attachment sources, the requirement entry binds the attachment source to the clause source that references the attachment. The resulting bidding requirement parsing set can be directly used as input for subsequent knowledge organization, knowledge retrieval, and bid document chapter generation. Compared to simply forming a keyword list, the bidding requirement parsing set simultaneously retains the requirement content, source location, and response location, enabling subsequent bid document generation to correspond to the chapter structure and table format of the bidding document.

[0037] S2. Using the bidding requirements parsing set, the bidding documents of enterprises are classified into chapters, recursively abstracted, and entity relationships are identified to construct a bidding knowledge tree index and a bidding knowledge graph index.

[0038] S2.1. Receive enterprise bidding documents and determine the source of enterprise bidding documents. Enterprise bidding documents include historical bids, technical solutions, qualification certificates, project cases, product information, engineering drawings and project video materials.

[0039] Among them, historical bids and technical solutions are derived from the company's existing bid documents and technical documents; qualification certificates are derived from scanned copies and electronic certificate files of the company's certificates; project cases are derived from the company's project acceptance materials and project introduction documents; product information is derived from product manuals, product pictures and parameter files; engineering drawings are derived from layout drawing files and scanned drawing files; and project video materials are derived from project site videos and project demonstration videos.

[0040] For each company's bidding documents, retain the document name, document type, page number, drawing number, video time period, and original file location to form a record of the source of the company's bidding documents.

[0041] Text content parsing is performed on the titles, paragraphs, tables, and appendix descriptions in historical bids and technical solutions from enterprise bidding documents to obtain textual knowledge content corresponding to enterprise capabilities, project experience, technical measures, personnel configuration, equipment configuration, implementation plan, and service commitments; text recognition is performed on the images and text in qualification certificates and product materials to obtain image and text knowledge content corresponding to certificate names, certificate numbers, validity periods, product names, and technical parameters; drawing text parsing is performed on drawing annotations, component names, and parameter descriptions in engineering drawings to obtain drawing annotation knowledge content; and content extraction is performed on video descriptions, subtitles, and key screen text in project video materials to obtain video description knowledge content.

[0042] Recording textual knowledge content, image-text knowledge content, drawing annotation knowledge content, and video instruction knowledge content as bidding knowledge items according to document name, document type, and original file location can prevent historical bid documents, certificates, drawings, and video materials from losing their source evidence during subsequent knowledge organization.

[0043] S2.2. Read the requirement categories, requirement content, chapter sources, clause sources, and table sources from the tender requirement analysis set. Assign the enterprise capability content from the enterprise bidding knowledge item to the qualification response chapter, the project experience content to the performance response chapter, the technical measures, equipment configuration, and implementation plan to the technical response chapter, the service commitments and delivery arrangements to the commercial response chapter, the product information and engineering drawings to the technical certification chapter, and the qualification certificates to the qualification certification chapter. When the same enterprise bidding knowledge item can correspond to multiple chapter uses, retain the original content of the enterprise bidding knowledge item and record the corresponding chapter use and requirement category in the chapter-based knowledge item.

[0044] The chapter-based knowledge entries are jointly determined by the enterprise bidding knowledge entries and the tender requirements analysis set. Compared with the method of simply classifying according to the enterprise's data folder name, the chapter-based knowledge entries can make the enterprise's bidding data directly correspond to the qualification requirements, technical requirements and commercial requirements of this tender document.

[0045] The chapter-based knowledge items are hierarchically divided through a recursive abstraction mechanism, generating bottom-level nodes, middle-level nodes, and top-level nodes.

[0046] The recursive abstraction mechanism first retains the original content of the chapter-based knowledge entries as the bottom-level nodes. The bottom-level nodes include the original knowledge content, document name, document type, original file location, and corresponding requirement category. Then, the bottom-level nodes are grouped according to project type, chapter purpose, technical theme, source of supporting materials, and service items. A summary is generated for the common content of the bottom-level nodes in the same group to obtain the middle-level nodes. The middle-level nodes include project case summaries, technical solution summaries, qualification certificate summaries, business commitment summaries, and service capability summaries. Finally, the enterprise capability content in the middle-level nodes that is relevant to this tender document is summarized at a higher level to obtain the top-level nodes, which are an overview of the enterprise capabilities.

[0047] The grouping fields used when grouping the underlying nodes include requirement category, chapter purpose, document type, project type, technical theme, source of supporting materials, service item, entity name, and document source location. Requirement category is used to distinguish between qualification requirements, performance requirements, personnel requirements, equipment requirements, pricing requirements, service response requirements, and format requirements. Chapter purpose is used to distinguish between qualification response chapters, performance response chapters, technical response chapters, commercial response chapters, technical certification chapters, and qualification certification chapters. Document type is used to distinguish between historical tenders, technical solutions, qualification certificates, project cases, product information, engineering drawings, and project video materials. Entity name corresponds to the qualification name, certificate name, project name, product name, technical parameter name, and service item name. Document source location retains the file name, page number, drawing number, video time period, and original file location.

[0048] The summary content of mid-level nodes retains the summary title, corresponding chapter purpose, corresponding requirement category, document name, document type, source location, entity name, original node path, and restrictive information fields. Restrictive information fields include qualification level, certificate number, issuing authority, validity period, project name, project owner, contract amount, service period, number of personnel, number of equipment, technical parameter values, technical parameter units, quotation amount, delivery deadline, and service response period. For qualification certificate summaries, the qualification level, certificate number, issuing authority, and validity period must not be deleted or generalized; for project case summaries, the project name, project owner, contract amount, service period, and acceptance status must not be deleted or generalized; for technical solution summaries, the technical parameter values, technical parameter units, equipment quantity, and implementation boundaries must not be deleted or generalized; for business commitment summaries, the quotation amount, delivery deadline, and service response period must not be deleted or generalized. Summary generation only merges and expresses non-restrictive descriptions, without changing the limiting conditions in the original knowledge content, enabling those skilled in the art to trace back to the corresponding original data using the retained fields.

[0049] It should be noted that the recursive abstraction processing mechanism retains the original data of the bottom-level nodes, and forms summary content of different granularities through the middle-level nodes and the top-level nodes. Compared with the ordinary vector knowledge base that only saves fragment text, the bidding knowledge tree index can support both original text-level retrieval and summary-level retrieval at the same time.

[0050] S2.3. Establish retrieval paths according to the summary hierarchy relationships between bottom-level nodes, middle-level nodes, and top-level nodes. Link each bottom-level node to its corresponding middle-level node, and link each middle-level node to the top-level node corresponding to the enterprise capability overview. Record the summary hierarchy relationships, chapter uses, corresponding requirement categories, and data source locations between the bottom-level nodes, middle-level nodes, and top-level nodes. When a bottom-level node belongs to multiple chapter uses, retain the original content of one bottom-level node in the bidding knowledge tree index and establish multiple retrieval paths for different middle-level nodes to avoid duplicate storage of the same enterprise bidding knowledge item.

[0051] Bidding entities are extracted from the bidding requirements analysis set and the bidding knowledge tree index to generate a bidding entity set. Bidding entities include project type, qualification name, technical parameters, product name, personnel position, certificate name, project case studies, and service items. Bidding-side entities representing the requirements of the bidding documents are extracted from the requirement category, requirement content, chapter source, clause source, and table source in the bidding requirements analysis set. These bidding-side entities include project type, qualification name, technical parameters, personnel position, and service items. Enterprise-side entities representing the enterprise's response content are extracted from the bottom-level, middle-level, and top-level nodes in the bidding knowledge tree index. These enterprise-side entities include product name, certificate name, project case studies, technical measures, and service commitments. The bidding-side entities and enterprise-side entities are recorded as a bidding entity set according to entity name, entity category, source location, and corresponding node path. The bidding entity set simultaneously covers both the bidding document requirements and the enterprise's information content, avoiding the situation where the bidding knowledge graph index only describes enterprise information and cannot correspond to the bidding requirements analysis set.

[0052] The bidding entity set is processed through name merging, hierarchical partitioning, and source location to generate an entity node set. Name merging addresses situations where the same qualification name, project case, product name, or service item appears repeatedly in the bidding materials of different companies, merging entities with identical or semantically equivalent text into the same entity node. Hierarchical partitioning addresses the hierarchical relationships between project type and project case, technical theme and technical parameters, qualification category and qualification name, and service item and service commitment, recording superior and subordinate entities in the same entity node path. Source location retains the position of the requirement item in the bidding requirement parsing set and the node path in the bidding knowledge tree index for each entity node.

[0053] Semantic equivalence is not determined solely by the output of the generative pre-trained language model, but rather through the enterprise's internal alias list, tender document terminology list, certificate number, project number, product model, or the same source location. When two entities have completely identical entity names, or entity names that match aliases in the enterprise's internal alias list, or at least one unique identifier among certificate number, project number, and product model, the two entities are merged into the same entity node. When there is only semantic similarity but no consistency criteria as described above, no merging of entities with the same name is performed, and they are only recorded as candidate related entity records.

[0054] S2.4. Determine the attribution, proof, response, and reference relationships among the entity node sets to form a bidding entity relationship chain. Before determining the relationships, verify the fields of qualification name, certificate name, project case, technical parameters, product name, personnel position, and service item in the entity node set; for qualification name entities, extract at least the qualification category, qualification level, issuing authority, validity period start and end dates, and source location; for certificate name entities, extract at least the certificate name, certificate number, issuing authority, certificate holder, validity period start and end dates, and source location; for project case entities, extract at least the project name, project owner, contract amount, service period, acceptance status, and source location; for technical parameter entities, extract at least the parameter name, parameter value, parameter unit, applicable product or equipment name, and source location; for product name entities, extract at least the product name, product model, key parameters, and product data source location; for personnel position entities, extract at least the position name, number of personnel, qualification certificate name, and source location; for service item entities, extract at least the service content, response period, service scope, and source location.

[0055] The attribution relationship is used to indicate that project cases belong to project types, technical parameters belong to technical topics, certificate names belong to qualification names, and service commitments belong to service items; the proof relationship is used to indicate that qualification certificates, project cases, product information, and engineering drawings can prove the company's capabilities, technical measures, and service capabilities; the response relationship is used to indicate that the company-side entity can respond to the requirement items in the tender requirements parsing set; the reference relationship is used to indicate the reference position of historical tender documents, technical solutions, qualification certificates, project cases, product information, engineering drawings, and project video materials in the tender document chapters.

[0056] The rules for determining attribution relationships are as follows: the project type, technical theme, qualification category, or service item fields of the subordinate entity are consistent with the corresponding fields of the superior entity, and the source location of the subordinate entity can be traced back to the bottom node in the bidding knowledge tree index; the rules for determining proof relationships are as follows: the qualification level, certificate number, project name, technical parameters, product model, or engineering drawing annotations in the proof material node can correspond to the capability content, technical measures, or service capabilities of the enterprise-side entity, and the proof material node has a file name, page number, drawing number, video time period, or original file location; the rules for determining response relationships are as follows: the entity category, entity name, or restrictive information fields of the enterprise-side entity can correspond to the requirement category, requirement content, and response location in the bidding requirement parsing set, and the restrictive information of the enterprise-side entity is not lower than the requirements of the bidding terms; the rules for determining citation relationships are as follows: the enterprise-side entity, proof material node, or project case node has a document name, document type, source location, and corresponding response location that can be written into the chapter of the bid, and corresponds to the chapter node in the bid generation path.

[0057] The bidding entity relationship chain uses a set of entity nodes as its foundation and connects these nodes with four types of relationships as edges, enabling the inclusion of bidding requirements, company information, and supporting documents within the same relationship network. When an entity node lacks supporting documents, a source location for information, or a node path in the bidding knowledge tree index, it is not output as a citationable source, does not participate in the output of proof relationships, and is only retained as a node to be supplemented for company information. When an entity node is semantically similar to the current response clause but lacks a response relationship, proof relationship, or citation relationship, it is only retained as a candidate related entity and is not output as supporting document knowledge.

[0058] S2.5. Bind the bidding entity relationship chain with the node paths in the bidding knowledge tree index to generate the bidding knowledge graph index.

[0059] For each bidding entity relationship chain, the corresponding bottom-level, middle-level, and top-level nodes in the bidding knowledge tree index are searched. The entity node, relationship type, node path, data source location, and corresponding requirement item are recorded together in the bidding knowledge graph index. When an entity node in a bidding entity relationship chain lacks a corresponding node path in the bidding knowledge tree index, the source location of the bidding requirement parsing set for the entity node is retained, and the entity node is used as the entity node for supplementing enterprise data, but it does not participate in the proof relationship output. When a bottom-level node in the bidding knowledge tree index does not have a corresponding requirement item in the bidding requirement parsing set, the bottom-level node is retained as a backup enterprise bidding knowledge item and is not included in the response relationship of the current bidding knowledge graph index. The resulting bidding knowledge graph index can be used together with the bidding knowledge tree index as input for subsequent multi-path recall retrieval. Compared to simply building a text retrieval index, the bidding knowledge graph index can express the proof relationship, response relationship, and reference relationship between entities, enabling subsequent candidate knowledge sets of clauses to retrieve enterprise data nodes, supporting material nodes, and project case nodes with clear sources.

[0060] S3. Call the bidding document tree, bidding knowledge tree index and bidding knowledge graph index to determine the clause category, find the relationship node, extract the summary and perform multi-way recall matching to generate a set of candidate clause knowledge.

[0061] S3.1. Select one requirement item from the tender requirement analysis set as the current response clause according to the chapter source and clause source. The current response clause includes the requirement category, requirement content, chapter source, clause source, table source, and response location.

[0062] The tender document tree is invoked to locate the chapter node and table source of the current response clause. Based on the parent and child nodes of the chapter node in the tender document tree, the semantic category of the current response clause is determined, and a clause category label is generated. The clause category label corresponds to one of the following categories: Qualification Response, Technical Response, Commercial Response, Quotation Description, Project Implementation, After-sales Service, and Deviation Description.

[0063] When a current response clause has both textual and table sources, the clause category label is determined by the chapter node, and the fill-in position is reserved by the table source location. This ensures that the current response clause corresponds to both the textual requirements and the table completion requirements in the tender document. Compared to directly classifying clauses by keywords, using the tender document tree to determine clause category labels reduces classification discrepancies caused by the repeated appearance of the same keywords in qualification clauses, technical clauses, and commercial clauses.

[0064] S3.2. Use clause category tags to retrieve the entity node set in the bidding knowledge graph index. First, search for the bidding-side entity node that matches the requirement category and requirement content of the current response clause in the entity node set. Then, search for the enterprise-side entity node that corresponds to the bidding-side entity node along the response relationship in the bidding knowledge graph index.

[0065] When there is a proof relationship between enterprise-side entity nodes, continue to search for the corresponding proof material nodes of qualification certificates, project cases, product information and engineering drawings connected to the enterprise-side entity nodes; when there is a reference relationship between enterprise-side entity nodes, continue to search for the reference positions of enterprise-side entity nodes in historical tender documents, technical solutions, qualification certificates, project cases, product information, engineering drawings and project video materials, and generate a clause relationship node chain.

[0066] S3.3. Utilize the clause relationship node chain to backtrack the top-level, middle-level, and bottom-level nodes in the bidding knowledge tree index, extract the enterprise capability overview, category summary, and original knowledge content, and generate a clause knowledge tree summary. Specifically, first, based on the enterprise-side entity nodes in the clause relationship node chain, search the bottom-level nodes in the bidding knowledge tree index to obtain the original knowledge content, document name, document type, and original file location; then, search the middle-level nodes along the retrieval path corresponding to the bottom-level nodes to obtain the corresponding category summaries from the project case summary, technical solution summary, qualification certificate summary, business commitment summary, and service capability summary; finally, search the top-level nodes to obtain the enterprise capability overview for this bidding document.

[0067] It should be noted that the knowledge tree summary of the clauses includes the original knowledge content, the classification summary, and the overall overview of the company's capabilities. This allows the subsequent main body of the tender document to both cite the original supporting materials and form a general expression that matches the current response clauses. Compared with the method of only recalling the original fragments, this can improve the organizational stability of the chapter content of long tender documents.

[0068] Starting from the tendering entity node corresponding to the current response clause, search for adjacent enterprise data nodes, supporting material nodes, and project case nodes along the proof relationship, response relationship, and citation relationship in the clause relationship node chain, and retain the relationship type, source location, and corresponding node path of adjacent nodes to generate the clause graph neighborhood.

[0069] The nodes in the clause graph neighborhood include the tendering entity node corresponding to the current response clause, the enterprise entity node that can respond to the current response clause, the supporting documentation node that can prove the enterprise entity node, and the reference node that can be used in the tender document section. The edges in the clause graph neighborhood include response relationships, proof relationships, and reference relationships. The clause graph neighborhood clarifies the connection path between tender document requirements, enterprise response content, supporting documentation, and reference locations, avoiding situations where multiple recall contents are semantically similar but lack supporting documentation or reference locations.

[0070] S3.4. Perform path compression and source sorting on the neighborhood of the clause graph to obtain the entity paths connected to the current response clause and the clause graph summary. The clause graph summary compresses entity connections to the available paths of the current response clause, which can reduce the entry of irrelevant enterprise information into the clause candidate knowledge set.

[0071] Among them, path compression refers to retaining the current response clause, enterprise-side entity node, supporting document node, and reference location node in the clause graph neighborhood, while deleting side nodes that have no response relationship, supporting document relationship, or reference relationship with the current response clause; source sorting refers to arranging the source positions in the entity path according to the clause source, table source, node path in the bidding knowledge tree index, and enterprise bidding data source record.

[0072] Based on the keywords, semantic content, and data type of the current response terms, keyword recall, semantic recall, and multimodal recall are performed respectively to obtain text matching knowledge, semantic matching knowledge, and supporting material knowledge.

[0073] Keyword recall uses the requirement category, clause name, qualification name, technical parameters, product name, personnel position, certificate name, project case and service items in the current response clause as query terms, and searches for content with direct text matching in enterprise bidding knowledge items, bottom nodes and middle nodes to obtain text matching knowledge; semantic recall uses the requirement content of the current response clause as query text, and uses vector retrieval to search for semantically similar historical solutions and project cases in the corresponding text content of historical bids, technical solutions and project cases to obtain semantic matching knowledge; multimodal recall is based on the document type of the current response clause, and searches for document content that can prove qualifications, parameters, equipment, project experience and service capabilities in qualification certificates, engineering drawings, product pictures and project video materials to obtain supporting material knowledge.

[0074] It should be noted that the direct correspondence between the text of the keyword recall guarantee clause, the semantic recall guarantee that historical solutions and project cases can be reused, and the multimodal recall guarantee that certificates, drawings, pictures and video materials can be included in the scope of supporting materials, together make up for the problem that single-path retrieval is prone to missing table items, supporting materials and historical cases.

[0075] S3.5. Arrange the clause knowledge tree summary, clause graph summary, text matching knowledge, semantic matching knowledge, and supporting material knowledge according to their chapter positions in the tender document tree to form a multi-path recall knowledge base. When arranging them, the clause knowledge tree summary is assigned to the chapter node corresponding to the current response clause, the clause graph summary is assigned to the entity path corresponding to the current response clause, the text matching knowledge is assigned to the clause source of the current response clause, the semantic matching knowledge is assigned to the requirement content of the current response clause, and the supporting material knowledge is assigned to the table source position and reference position of the current response clause.

[0076] The multi-path recall knowledge base retains the source, purpose, and correspondence of each recall content, enabling the subsequent generation of the main text of the tender document to distinguish between the content that can be written in the main text, the content that can be cited in the supporting materials, and the content that can be filled in the tables. This is different from the problem of ordinary recall results being arranged only according to similarity and unable to correspond to the chapter position in the tender document tree.

[0077] The knowledge content that matches the clause category label, table source location and entity path is selected from the multi-path recall knowledge base to generate a clause candidate knowledge set. When selecting, the clause knowledge tree summary and text matching knowledge that match the clause category label are retained, the supporting material knowledge that matches the table source location is retained, and the clause graph summary and semantic matching knowledge that match the entity path are retained.

[0078] When selecting knowledge content, it is prioritized and filtered according to the priority of multi-path recall knowledge: content with proof, response, and reference relationships is given priority over content with only one type of relationship; content with multiple source fields such as file name, page number, clause number, table row and column position, and attachment number is given priority over content with incomplete source fields; knowledge of supporting materials is given priority over knowledge of simple semantic matching; for multiple qualification certificates corresponding to the same qualification name or the same certificate name, the qualification certificate whose validity period covers the bid closing date in the bidding documents is used as the available certificate. When multiple available certificates exist at the same time, the qualification certificate with the later expiration date is given priority. If the bidding documents do not parse to the bid closing date, the bid generation date is used as the validity comparison benchmark and recorded in the source position; content that is only semantically similar but does not have proof, response, or reference relationships can only be used as alternative content and cannot be used as supporting material knowledge output.

[0079] When duplicate knowledge content exists in the multi-path recall knowledge base, the knowledge content with the more complete source location is retained. When knowledge content in the multi-path recall knowledge base is only semantically similar but lacks proof or citation relationships, semantically matched knowledge is retained as an alternative and not used as proof material knowledge. When proof material knowledge and semantically matched knowledge correspond to the same response clause but their content is inconsistent, the proof material knowledge with proof, response, or citation relationships and complete source fields is used as the citationable source, and semantically matched knowledge is only used to assist in paragraph expression and is not used for citation of proof material.

[0080] S4. Based on the candidate knowledge set of clauses, intermediate results of tables and format constraints, select chapter paths in the mind tree, and trigger forward search enhancement before paragraph generation to generate the first draft of the tender.

[0081] S4.1. Determine the format constraints based on the tender requirements analysis set and intermediate table results. The format constraints include chapter order, heading level, table items, response scope, and the location of supporting documents references. The chapter order is derived from the arrangement of parent and child nodes in the tender document tree. The heading level is derived from the table of contents and clause number in the tender document. The table items are derived from the intermediate table results corresponding to the qualification review form, scoring form, technical parameter form, quotation list, and service commitment form. The response scope is derived from the requirement category and requirement content in the requirement items. The location of supporting documents references is derived from the supporting document knowledge, clause graph summary, and entity path in the clause candidate knowledge set.

[0082] Format constraints unify the chapter structure, table entry locations, and sources of supporting materials in the bidding documents within the same generation boundary. Compared to simply generating the main text according to a general bid template, this avoids the disconnect between chapter order, table items, and the location of supporting material citations.

[0083] The candidate knowledge set of clauses is used to match the chapter nodes in the bidding document tree to establish chapter candidate paths.

[0084] Specifically, the clause category tags in the clause candidate knowledge set are matched with the corresponding nodes in the tender document tree, including project overview, qualification requirements, technical requirements, commercial requirements, pricing requirements, delivery requirements, and review rules. Qualification response, technical response, commercial response, pricing description, project implementation, after-sales service, and deviation description are then used as candidate chapter directions. When a clause candidate knowledge set corresponds to multiple chapter nodes, chapter nodes consistent with the clause category tags are prioritized, while chapter nodes consistent with the table source location are also retained as candidate branches, forming a chapter candidate path. The chapter candidate path retains the current response clause, chapter node, table source location, and candidate chapter direction, ensuring that subsequent thought tree reasoning revolves around the actual tender document structure rather than a fixed chapter template.

[0085] S4.2. Combining the intermediate results of the tables and format constraints, verify the location of table items, supporting materials, and response criteria in the candidate chapter paths to generate a set of optional chapter paths. The table item location verification determines whether the candidate chapter paths can accommodate the items to be filled in the qualification review form, scoring form, technical parameter table, quotation list, and service commitment form; the supporting material location verification determines whether the supporting material knowledge in the candidate clause knowledge set has corresponding source locations such as qualification certificates, project cases, product information, engineering drawings, and project video materials; the response criteria verification determines whether the expression direction of the candidate chapter paths is consistent with the requirement category, requirement content, and clause category label.

[0086] Through the above verification, candidate chapter paths that can simultaneously satisfy the requirements of chapter node, table item location, supporting material location, and response caliber are retained, and a set of optional chapter paths is generated.

[0087] The mind tree reasoning process is used to deduce the clause continuation relationship, material reference relationship and paragraph order of the optional chapter path set, and generate chapter path deduction results.

[0088] Specifically, each candidate chapter path in the optional chapter path set is first treated as a candidate branch, and the candidate chapter direction, chapter node, current response clause, table source location, and clause category label are written in the candidate branch. Then, according to the parent and child nodes in the bidding document tree, the current response clauses under the same candidate branch are sequentially attached to the corresponding chapter nodes to form a clause arrangement sequence. Subsequently, according to the supporting material knowledge, clause diagram summary, and entity path in the clause candidate knowledge set, the supporting material citation locations corresponding to qualification certificates, project cases, product information, engineering drawings, and project video materials are attached to the corresponding current response clauses in the clause arrangement sequence to form a material citation sequence. Finally, according to the heading level and table items in the format constraints, the clause arrangement sequence and material citation sequence are arranged into paragraph order to obtain the chapter path deduction result.

[0089] The chapter path derivation results include chapter arrangement, clause order, table item location, material reference location, and paragraph sequence. Each candidate branch generates a chapter path derivation result according to the above clause attachment, material attachment, and paragraph positioning process. Compared to directly generating a single path, the above process can compare the consistency between different chapter organization methods and the bidding document tree, clause candidate knowledge set, and format constraints before the initial draft of the tender document is generated.

[0090] S4.3. Select the chapter path derivation results that are consistent with the bidding document tree, the candidate knowledge set of clauses, and the format constraints from the chapter path derivation results, and generate the bid document generation path. When selecting, check whether the chapter path derivation results conform to the chapter order in the bidding document tree, check whether the chapter path derivation results cover the current response clause, clause knowledge tree summary, clause graph summary, text matching knowledge, semantic matching knowledge, and supporting material knowledge in the candidate knowledge set of clauses, and check whether the chapter path derivation results meet the heading level, table items, response scope, and supporting material citation position in the format constraints. When multiple chapter path derivation results meet the above conditions, give priority to the chapter path derivation results with complete source location, clear supporting material citation position, and paragraph order consistent with the bidding document tree.

[0091] For each chapter's path derivation result, a path priority record is established. This record includes items related to chapter order consistency, clause coverage completeness, table item location correspondence, supporting document citation location existence, and response consistency. Chapter order consistency indicates whether the chapter arrangement conforms to the parent and child node order in the tender document tree. Clause coverage completeness indicates whether all current response clauses are attached to their corresponding chapter nodes. Table item location correspondence indicates whether the table items in the qualification review form, scoring form, technical parameter form, quotation list, and service commitment form correspond to the response clauses. Supporting document citation location existence indicates the presence of citations in qualification certificates, project cases, product information, engineering drawings, and project video materials. Response consistency indicates whether the paragraph's expression direction is consistent with the requirement category, requirement content, and clause category label. If any current response clause in the chapter path derivation result lacks a corresponding chapter node, or if the response consistency is inconsistent with the requirement content, then that chapter's path derivation result is not considered an optional generated path.

[0092] When multiple chapter path derivation results all satisfy the criteria of consistent chapter order, complete clause coverage, corresponding table items, existence of supporting material citations, and consistent response, the final chapter path derivation result is selected in the following order: completeness of supporting material citations > completeness of clause coverage > consistency of chapter order > consistency of table items. The completeness of supporting material citations is prioritized to ensure that each citationable paragraph has a source of supporting material; the completeness of clause coverage is used to ensure that all current response clauses in the tender requirements analysis set are responded to; the consistency of chapter order is used to ensure that the chapters in the tender document are consistent with the tender document tree; and the consistency of table items is used to ensure that the paragraphs in the main text correspond to the table filling positions.

[0093] The tender document generation path is derived from the selected chapter path. This path includes chapter arrangement, clause order, table item location, material citation location, and paragraph sequence. When determining the current chapter to be generated and the next paragraph location using the tender document generation path, first, the first chapter before the main text is generated is selected as the current chapter to be generated, according to the chapter arrangement in the path. Then, according to the clause order and paragraph sequence within the current chapter to be generated, the first current response clause or table item before the main text is generated is selected as the next paragraph location. When the next paragraph location corresponds to a table item, the table item location and the current response clause are used together as the source of the next paragraph location. When the next paragraph location corresponds to a supporting material citation location, the supporting material citation location and the current response clause are used together as the source of the next paragraph location.

[0094] The proposed paragraph semantics are formed based on the current chapter to be generated and the position of the next paragraph. The proposed paragraph semantics are formed based on the chapter title of the current chapter to be generated, the current response clause corresponding to the next paragraph position, the clause knowledge tree summary in the clause candidate knowledge set, and the response criteria in the format constraints. The proposed paragraph semantics are used to describe the topic, response matters, direction of supporting material citation, and paragraph order of the proposed next paragraph. Clause names, technical matters, qualification matters, case matters, and supporting material names are extracted from the proposed paragraph semantics to generate the paragraph search intent. The paragraph search intent can limit the search scope before the formal paragraph is generated, which reduces the risk of insufficient source support for the main text content compared to generating the main text first and then searching for sources.

[0095] S4.4. Based on the paragraph retrieval intent, the candidate knowledge set of clauses, the bidding knowledge tree index, and the bidding knowledge graph index are retrieved to obtain the source content corresponding to the next paragraph position, generating a paragraph source knowledge fragment. When retrieving the candidate knowledge set of clauses, text matching knowledge, semantic matching knowledge, and supporting material knowledge corresponding to the paragraph retrieval intent are obtained; when retrieving the bidding knowledge tree index, the original knowledge content in the bottom-level nodes, the classification summary in the middle-level nodes, and the enterprise capability overview in the top-level nodes are obtained along the retrieval path corresponding to the clause knowledge tree summary; when retrieving the bidding knowledge graph index, the enterprise-side entity nodes, supporting material nodes, and citation location nodes are obtained along the entity path corresponding to the clause graph summary. The paragraph source knowledge fragment retains the document name, document type, source location, corresponding clause name, and supporting material citation location, enabling the paragraph generation process to return to the source of the enterprise's bidding documents.

[0096] The paragraph generation prompts are designed by arranging the semantics of the paragraphs, the knowledge fragments from which they originate, format constraints, and the path to generate the tender document. They also define the chapter titles, response items, cited materials, and paragraph order. The chapter titles in the prompts are derived from the tender document generation path, the response items from the semantics of the paragraphs, the cited materials from the knowledge fragments from which they originate, and the paragraph order from the format constraints. The prompts further specify that the qualification response section should cite company qualifications, certificates, and performance records; the technical response section should cite technical solutions, implementation measures, product information, and project case studies; the commercial response section should cite service commitments, delivery arrangements, and pricing specifications; and the deviation explanation section should be consistent with the tender terms and the company's response. By fixing "what to write, what to cite, and where to place" before paragraph generation, the prompts reduce inconsistencies between the main body of the tender document and the tender terms, supporting documents, and forms.

[0097] S4.5. When the knowledge fragment from which the paragraph originates is consistent with the semantics of the paragraph, a formal paragraph is generated according to the paragraph generation prompts. Specifically, the generative pre-trained language model determines the paragraph position based on the chapter title in the paragraph generation prompts, determines the paragraph topic based on the response items, writes the corresponding content into the enterprise qualifications, certificates, performance, technical solutions, implementation measures, product information, project cases, service commitments, delivery arrangements, and quotation descriptions based on the cited materials, determines the paragraph's position in the current chapter to be generated based on the paragraph order, and generates the formal paragraph.

[0098] When a paragraph's source knowledge fragment lacks corresponding clause content, the candidate knowledge set is invoked again to supplement the source content and correct the paragraph's semantics, generating a corrected paragraph. Missing clause content includes the absence of clause name content, supporting document citation locations, table item locations, or company information consistent with the response guidelines in the paragraph's source knowledge fragment. When invoking the candidate knowledge set, priority is given to supplementing clause knowledge tree summaries and text matching knowledge consistent with clause category tags, followed by clause graph summaries and semantic matching knowledge consistent with entity paths, and finally, supporting document knowledge consistent with table source locations. The corrected paragraph's semantics and the supplemented source content are then rearranged into paragraph generation prompts to generate the corrected paragraph. Through missing content callbacks and paragraph semantic correction, paragraph generation avoids situations where supporting documents, table references, or clause sources are missing.

[0099] The initial draft of the tender document is generated by arranging the formal and revised paragraphs according to the tender document generation path. During the arrangement, the qualification response, technical response, commercial response, quotation explanation, project implementation, after-sales service, and deviation explanation are first arranged according to the chapter order in the tender document tree. Then, the chapter titles and subheadings are arranged according to the heading hierarchy in the format constraints. Following this, the table response content is written according to the position of the table items in the intermediate table results, and the reference positions of supporting materials are preserved in the corresponding paragraphs. The initial draft of the tender document maintains the correspondence between chapter titles, clause order, response scope, table items, and sources of supporting materials, serving as input for proofreading and output. Compared to template-filled text, the initial draft of the tender document generated in this way can dynamically form chapter content based on the current chapter structure of the tender document, the set of candidate clauses, and the sources of supporting materials, reducing problems such as chapter misalignment, missing materials, and inconsistent response scope.

[0100] S5. Verify the consistency of sources and revise the initial draft of the tender document, the set of tender requirements parsing, and the set of candidate clauses through natural language reasoning chain, and generate the final draft of the tender document.

[0101] S5.1. Extract the qualification response, technical response, commercial response, quotation explanation, project implementation, and after-sales service content according to the chapter titles and paragraph positions in the initial draft of the tender document; segment the qualification response, technical response, commercial response, quotation explanation, project implementation, and after-sales service content into response sentences to form a sequence of response sentences. When segmenting response sentences, use periods, semicolons, line breaks, numbering levels, and table cell boundaries as the segmentation criteria, and retain the chapter title, paragraph position, clause name, and supporting document citation location of each response sentence.

[0102] When a paragraph contains responses to multiple tender terms, it is split into multiple response sentences according to the term name and response item. The resulting sequence of response sentences can transform long paragraphs in the initial draft of the tender into text units that can be verified sentence by sentence. Compared to proofreading only the entire text, this method can pinpoint the problem locations in specific qualification responses, technical responses, commercial responses, quotation explanations, project implementation, and after-sales service content.

[0103] S5.2. Through natural language reasoning chain, the consistency of the response sentence sequence with the bidding terms, source content, knowledge tree summary and graph summary is judged, and the source consistency verification result is generated.

[0104] The tender terms are derived from the requirements, chapter sources, clause sources, and table sources in the tender requirements analysis set; the source content comes from the text matching knowledge, semantic matching knowledge, and supporting material knowledge in the clause candidate knowledge set; the knowledge tree summary comes from the clause knowledge tree summary in the clause candidate knowledge set; and the graph summary comes from the clause graph summary in the clause candidate knowledge set.

[0105] The Natural Language Inference Chain uses the response sentence as the text to be verified and the bidding terms, source content, knowledge tree summary, and graph summary as the verification basis. It determines whether the response sentence is supported by the source content, meets the requirements of the bidding terms, is consistent with the knowledge tree summary, and is consistent with the entity paths and proof relationships in the graph summary. The source consistency verification result includes the position of the response sentence, the corresponding bidding terms, the corresponding source content, the verification status, and the problem type; the verification status includes source support, lack of source support, and content discrepancies. Compared to proofreading methods that only check for typos and formatting errors, the Natural Language Inference Chain can establish a correspondence between each response sentence in the initial draft of the bid and the bidding requirements and the company's data sources, reducing the risk of unfounded commitments, incorrect citations of materials, and deviations in the response to terms.

[0106] Among them, "source support" means that the clause name, requirement category, response item, restrictive information field, and supporting material reference location of the response sentence can all correspond to the bidding clause, source content, knowledge tree summary, and graph summary respectively, and the response sentence does not change the qualification level, validity period, project name, amount, quantity, technical parameter value, delivery period, and service response period; "lack of source support" means that the response sentence lacks at least one necessary source among corresponding text matching knowledge, semantic matching knowledge, supporting material knowledge, knowledge tree summary, or graph summary, or there is no proof relationship, response relationship, or reference relationship in the entity path corresponding to the response sentence; "content difference" means that the qualification requirements, performance requirements, personnel requirements, equipment requirements, quotation requirements, service response requirements, format requirements, supporting material reference location, or restrictive information field in the response sentence are inconsistent with the bidding clause, source content, knowledge tree summary, or graph summary.

[0107] When the same response sentence simultaneously meets the criteria for multiple verification statuses, the final verification status is determined in the following order: content difference takes precedence over lack of source support, and lack of source support takes precedence over source support. When there are conflicts between source contents, the knowledge of supporting materials with proof relationships, response relationships, and citation relationships and complete source fields takes precedence, followed by the summary of the clause knowledge tree and the summary of the clause diagram, and then the knowledge of text matching. Semantic matching knowledge is only used as an auxiliary basis for judgment. When there are conflicts between old and new versions of qualification certificates, project cases, technical parameters, or service commitments, the supporting materials that can cover the bid deadline and have a later expiration date take precedence. Supporting materials that cannot cover the bid deadline are not considered as source support. When the response sentence is over-responding, under-responding, or the cited materials do not match the bidding terms, the verification status is determined to be content difference.

[0108] When the source consistency verification result indicates that the response sentence lacks source support, the candidate knowledge set of clauses is invoked to supplement the corresponding source content and generate a supplementary response sentence. Missing source support includes unmatched text knowledge, unmatched semantic knowledge, unmatched supporting material knowledge, unmatched knowledge tree summary, and unmatched entity paths in the graph summary. For response sentences lacking source support, the clause category tag in the candidate knowledge set of clauses is first searched based on the chapter title and clause name of the response sentence. Then, text matching knowledge and clause knowledge tree summary consistent with the clause category tag are obtained from the candidate knowledge set of clauses. Subsequently, clause graph summary and semantic matching knowledge consistent with the entity path are obtained. Finally, supporting material knowledge consistent with the source location of the table and the reference location of the supporting material is obtained. The supplementary response sentence retains the original response sentence's theme and chapter position while incorporating supplementary source content, and retains the document name, document type, source location, and supporting material reference location. By using supplementary response sentences, text content lacking source support can be transformed into response content with enterprise data support, reducing the omission of supporting materials and the loss of source locations.

[0109] When the source consistency verification result indicates that there is a content difference between the response sentence and the tender terms, the corresponding response sentence is rewritten and a revised response sentence is generated by calling the tender requirements parsing set and the clause candidate knowledge set.

[0110] The content differences include inconsistencies between the qualification requirements, performance requirements, personnel requirements, equipment requirements, quotation requirements, service response requirements, and format requirements in the response sentences and the requirements in the tender requirements analysis set, as well as inconsistencies between the location of the supporting documents cited in the response sentences and the supporting documents knowledge in the candidate knowledge set of clauses.

[0111] For response sentences with content differences, first obtain the requirement category, requirement content, chapter source, clause source, and table source of the corresponding requirement item from the tender requirement parsing set. Then, obtain the source content that is consistent with the requirement category, table source location, and entity path from the clause candidate knowledge set. Finally, rewrite the response sentence according to the response guidelines in the tender clause to generate a revised response sentence.

[0112] The revised response sentences retain their original section positions, and inconsistent expressions are replaced with expressions consistent with the tender terms and source content. Revising the response sentences can resolve issues of over-response, under-response, and mismatched cited materials in the initial draft of the tender.

[0113] S5.3. Receive user dialogue content submitted via a dialog and use this content to locate the target chapters in the initial draft of the tender document. User dialogue content includes chapter names, clause names, modification semantics, replacement requirements, expansion requirements, and compression requirements.

[0114] When user dialogue contains chapter names, the target chapter in the initial draft of the tender is matched according to the chapter name; when user dialogue contains clause names, the position of the response sentence in the response sentence sequence is matched according to the clause name; when user dialogue only contains modification semantics, the modification semantics are semantically matched with the chapter titles, response items, and clause names in the initial draft of the tender to determine the target chapter and target response sentence. Locating the target chapter through user dialogue content can limit the manual modification intention to a specific chapter and a specific response sentence, avoiding indiscriminate rewriting of the initial draft of the tender that would cause changes to the chapter structure.

[0115] The target chapter is revised by combining supplementary and revised response sentences to generate the final tender document. When revising the target chapter, if a response sentence in the target chapter indicates a lack of supporting sources in the source consistency verification result, a supplementary response sentence is added to the corresponding position in the target chapter. If a response sentence in the target chapter indicates a difference between the source consistency verification result and the tender terms, a revised response sentence replaces the corresponding position in the target chapter. If the user dialogue includes a replacement request, the corresponding response sentence in the target chapter is replaced according to the semantic changes in the user dialogue. If the user dialogue includes an expansion request, source content consistent with the candidate knowledge set of the terms is added without changing the tender terms, source content, knowledge tree summary, and graph summary. If the user dialogue includes a compression request, the tender terms response content, supporting material citation positions, and source positions in the target chapter are retained, and duplicate expressions are deleted.

[0116] After the user dialogue is revised, the source consistency verification is re-executed on the revised target chapter or target response sentence. During the re-verification, the source content in the bidding requirements analysis set, the source content in the candidate knowledge set, the knowledge tree summary, and the graph summary are still used as the verification basis. If the re-verification result is source support, the revised target chapter or target response sentence is written into the final bid document. If the re-verification result is that source support is lacking, the candidate knowledge set is called to supplement the source content, and a supplementary response sentence is generated again based on the supplemented source content and written into the final bid document. If the re-verification result is that the content is different, the bidding requirements analysis set and the candidate knowledge set are called to rewrite the corresponding response sentence, and the rewritten response sentence is written into the final bid document after it meets the source support. If the source support still cannot be achieved after supplementation or rewriting, the corresponding content is marked as content to be confirmed, and the target chapter position, question type, missing source field, and user dialogue content are retained for manual confirmation.

[0117] After the content revision is completed, all chapters are arranged according to the chapter arrangement, clause order, table item location, material reference location, and paragraph sequence in the tender document generation path to generate the final tender document. The final tender document maintains the correspondence between the tender clauses, source content, knowledge tree summary, graph summary, user dialogue content, and main text paragraphs. Compared to the method of simply reviewing and uniformly modifying by humans, it can complete the dialogue revision while preserving traceable knowledge sources, reducing chapter structure disruption, source loss, and deviation of response caliber.

[0118] In summary, this invention does not simply generate bid text using a large model. Instead, it first uses the intermediate results of the bidding document tree and tables to uniformly parse the main text clauses, table items, attachment descriptions, and their source locations into a bidding requirement parsing set. This parsing set then drives the formation of a bidding knowledge tree index and a bidding knowledge graph index from the enterprise's bidding data, enabling the enterprise's data to be organized and accessed according to the relationship of "bidding requirements - enterprise response content - supporting materials - citation location". Compared to existing bid document generation methods that rely on template filling, keyword retrieval, or single-vector recall, this invention prioritizes knowledge content with proof relationships, response relationships, citation relationships, and complete source fields through a multi-path recall knowledge base. Before paragraph generation, it uses mind tree chapter path selection and forward retrieval to enhance the determination of chapter order, response scope, table location, and supporting material citation location, reducing unfounded generation, miscitation of materials, and omission of clauses from the source. Furthermore, by using natural language reasoning chains to verify the consistency of sources between the initial draft of the tender and the revised content of the dialogue, the final tender is made consistent in terms of clause response, knowledge sources, supporting materials and chapter structure, thereby improving the traceability, accuracy and review adaptability of the automatically generated tender.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base, characterized in that: include, The content of the chapters and tables in the bidding documents are broken down hierarchically to construct a bidding document tree and intermediate table results, and bidding elements are extracted to generate a set of bidding requirements analysis. The bidding requirements parsing set is used to classify, recursively abstract, and identify entity relationships in the enterprise bidding materials, and to construct a bidding knowledge tree index and a bidding knowledge graph index. The tender document tree, tender knowledge tree index, and tender knowledge graph index are used to determine clause categories, find relationship nodes, extract summaries, and perform multi-way recall matching to generate a set of candidate clause knowledge. Based on the candidate knowledge set of clauses, intermediate results in tables, and format constraints, the mind tree chapter path is selected, and a forward search enhancement is triggered before paragraph generation to generate the first draft of the tender document; By using a natural language reasoning chain, the initial draft of the tender document, the set of tender requirements parsing, and the set of candidate clauses knowledge are verified for consistency and revised through dialogue to generate the final draft of the tender document.

2. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 1, characterized in that: The specific steps for constructing the tender document tree and intermediate table results are as follows: The layout of the bidding documents, including the table of contents, clause numbers, main text paragraphs, and appendix descriptions, is analyzed. A generative pre-trained language model is used to determine the semantic category and generate a bidding document tree. The system identifies row and column positions, completes semantic information for rows and columns, assigns items to their correct positions, and connects content across cells in the qualification review form, scoring form, technical parameter form, quotation list, and service commitment form in the bidding documents, generating intermediate table results.

3. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 2, characterized in that: The specific steps for generating the tender requirement parsing set are as follows: Based on the bidding document tree and intermediate table results, extract the chapter bidding item set and table bidding item set respectively, and arrange the chapter bidding item set and table bidding item set into a bidding element chain according to the chapter source, clause source and table source; Establish requirement entries based on the source location and item category in the bidding element chain, and generate a bidding requirement parsing set.

4. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 3, characterized in that: The specific steps for constructing the bidding knowledge tree index are as follows: The content of enterprise bidding documents is analyzed, and the purpose of each chapter is divided into chapters using the bidding requirements analysis set, generating chapter-based knowledge items; By using a recursive abstraction processing mechanism, chapter-based knowledge items are hierarchically divided, so that the bottom-level nodes retain the original knowledge content, the middle-level nodes form a classification summary, and the top-level nodes form an overview of the enterprise's capabilities. The retrieval path is established based on the summary hierarchy among the bottom-level nodes, middle-level nodes, and top-level nodes, and a bidding knowledge tree index is generated.

5. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 4, characterized in that: The specific steps for constructing the bidding knowledge graph index are as follows: The bidding entities are extracted from the bidding requirements parsing set and the bidding knowledge tree index, and the bidding entities are merged with the same name, divided into hierarchical and source locations to generate an entity node set; Determine ownership, proof, response, and reference relationships among entity node sets, and bind the corresponding relationships to node paths in the bidding knowledge tree index to generate a bidding knowledge graph index.

6. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 5, characterized in that: The specific steps for generating the candidate knowledge set of terms are as follows: The current requirement item is selected as the current response clause from the tender requirement parsing set. The chapter node and table source position of the current response clause are located by calling the tender document tree, the semantic category of the current response clause is determined, and a clause category label is generated. The entity node set in the bidding knowledge graph index is retrieved by using clause category tags. Enterprise data nodes that have proof relationships, response relationships, and reference relationships with the current response clause are filtered to generate a clause relationship node chain. By tracing back the top-level, middle-level, and bottom-level nodes in the bidding knowledge tree index using the clause relationship node chain, the enterprise capability overview, category summary, and original knowledge content are extracted to generate a clause knowledge tree summary. Starting from the entity node corresponding to the current response clause, search for adjacent enterprise data nodes, supporting material nodes, and project case nodes along the proof relationship, response relationship, and citation relationship in the clause relationship node chain, and retain the relationship type and source location between adjacent nodes to generate the clause graph neighborhood; Path compression and source sorting are performed on the neighborhood of the clause graph to obtain the entity path and clause graph summary connected to the current response clause; Based on the keywords, semantic content, and data type of the current response terms, keyword recall, semantic recall, and multimodal recall are performed respectively to obtain text matching knowledge, semantic matching knowledge, and supporting material knowledge; The knowledge tree summary of clauses, the graph summary of clauses, the text matching knowledge, the semantic matching knowledge, and the supporting material knowledge are arranged according to the chapter positions in the bidding document tree to form a multi-path recall knowledge base; Select knowledge content from the multi-path recall knowledge base that matches the clause category label, table source location, and entity path to generate a clause candidate knowledge set.

7. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 1, characterized in that: The format constraints include the chapter order, heading level, table items, response criteria, and the location of supporting documents as determined by the tender requirements parsing set and intermediate table results.

8. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 7, characterized in that: The specific steps for selecting chapter paths in the mind tree are as follows: The candidate knowledge set of clauses is used to match the chapter nodes in the bidding document tree to establish the corresponding chapter candidate paths. The table intermediate results and format constraints are combined to verify the position of table items, the position of supporting materials and the response scope in the chapter candidate paths, and generate a set of optional chapter paths. The tender document generation path is generated by using a mind tree reasoning process to deduce the clause continuation relationship, material reference relationship and paragraph order of the optional chapter path set, and selecting the chapter path deduction result that is consistent with the tender document tree, the candidate knowledge set of clauses and format constraints.

9. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 8, characterized in that: The specific steps for triggering lookup retrieval enhancement before paragraph generation are as follows: The tender document generation path is used to determine the current chapter to be generated and the position of the next paragraph, forming the semantics of the paragraph writing, and identifying clause names, technical matters, qualification matters, case matters and supporting material names to generate paragraph retrieval intent; Based on the candidate knowledge set of paragraph retrieval intent callback clauses, the bidding knowledge tree index, and the bidding knowledge graph index, the source content corresponding to the next paragraph position is obtained, and a paragraph source knowledge fragment is generated. The semantics of paragraph writing, knowledge fragments from paragraph sources, format constraints, and the path to generate the tender document are arranged into paragraph generation prompts, and the chapter titles, response items, cited materials, and paragraph order are limited. When the source knowledge fragment of a paragraph is consistent with the semantics of the paragraph, a formal paragraph is generated. When the source knowledge fragment of a paragraph lacks corresponding content of the clause, the candidate knowledge set of the clause is called again to supplement the source content and correct the semantics of the paragraph, and a revised paragraph is generated. Arrange the formal and revised paragraphs according to the tender document generation path to generate the initial draft of the tender document.

10. The method for automatically generating tender documents based on a large model combined with a multi-path recall knowledge base as described in claim 9, characterized in that: The specific steps for generating the final draft of the tender document are as follows: The initial draft of the tender document is divided into response sentences for the qualification response, technical response, commercial response, quotation explanation, project implementation and after-sales service content, forming a sequence of response sentences; The natural language reasoning chain is used to determine the consistency between the response sentence sequence and the bidding terms, source content, knowledge tree summary and graph summary, and generate source consistency verification results. When the source consistency verification result indicates that the response sentence lacks source support, the candidate knowledge set of the clauses is invoked to supplement the corresponding source content and generate a supplementary response sentence. When the source consistency verification result indicates that there is a content difference between the response sentence and the tender terms, the corresponding response sentence is rewritten by calling the tender requirements parsing set and the clause candidate knowledge set, and a revised response sentence is generated. The target chapters in the initial draft of the tender document are located using user dialogue content, and the content of the target chapters is revised by combining supplementary response sentences and revision response sentences to generate the final draft of the tender document.