CN-122019481-A - Automatic bidding document generation method based on large language model and workflow arrangement

CN122019481ACN 122019481 ACN122019481 ACN 122019481ACN-122019481-A

Abstract

The invention discloses a bid file automatic generation method based on a large language model and workflow arrangement, which automatically extracts keyword contents such as a bid file format, a negotiation response book format, a competitive negotiation response file format and the like from a bid file, and identifies a standard catalog required by the bid file. And automatically constructing a subdirectory for the "technical file" section based on each score listed in the scoring criteria. And (3) introducing scoring detail semantic analysis, automatically generating each sub-chapter title through keyword matching and rule extraction, ensuring that the technical file content corresponds to scoring points one by one, and improving the scoring rate. And (3) checking the integrity and the accuracy of the automatically generated directory structure, and if the directory structure is incompletely identified or is disordered logically, returning the flow, re-extracting or adjusting the system to ensure that the directory accurately meets the requirement of the bidding documents.

Inventors

PENG LING
LIU WENFENG
LIU XINYI
JIN LIANG

Assignees

江西飞尚科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (7)

1. A bid file automatic generation method based on large language model and workflow arrangement is characterized in that the bid file automatic generation method based on large language model and workflow arrangement comprises the following specific steps: Step 1, uploading a bid-inviting file by a user, uploading the project bid-inviting file by the user through a front-end interface, supporting PDF, word, TXT file formats, and automatically identifying the type of a document by a system and analyzing the document; Step 2, extracting and generating a directory structure, positioning a section of a bidding file format by a system, extracting directory structure items in the section, identifying a section of a scoring standard, extracting all clear scoring points as subdirectories of a technical file, integrating and generating a complete bidding file directory tree, wherein the structure covers business response, technical response and service guarantee core parts; step 3, checking and repairing the catalogue, comparing the bidding documents, and executing logic integrity and structure normalization check; Step4, constructing a prompting word and generating chapter contents, constructing a structured prompting word template aiming at each catalog node, and inputting the template into a large language model to generate response contents; Step 5, integrating the knowledge base with external search content, calling the enterprise knowledge base, extracting related chapters of past bidding file content filling of enterprises, calling hundred-degree search nodes, obtaining latest industry specifications, standard terms or reference text, structurally integrating search results into prompt words through workflow arrangement, and improving rationality and coverage of generated content; and 6, expanding and coloring the chapter content, wherein for chapters which do not meet the word number or the content list, the system can automatically expand the content according to a set threshold value.
2. The automatic generation method of bidding documents based on large language model and workflow layout according to claim 1, wherein in step1, the user can set generation parameters, view catalog and chapter content and conduct content fine adjustment through the front-end interface.
3. The automatic bidding document generation method based on large language model and workflow layout according to claim 1, wherein the front-end interface adopts a flow layout engine for controlling the task flow of each step, and has the capabilities of condition judgment, error rollback and branch logic control.
4. The method for automatically generating a bid document based on a large language model and workflow layout of claim 1 wherein in step 4, constructing a structured prompt word template comprises chapter titles, score points, project backgrounds, industry terms.
5. The automatic generation method of bidding documents based on large language model and workflow layout according to claim 1, wherein in step 5, the large language model is called to generate chapter content, and enterprise knowledge base and hundred-degree search result are integrated to enrich content semantics.
6. The automatic generation method of bidding documents based on large language model and workflow layout according to claim 1, wherein in step 5, the generated catalogue and chapter contents need to be checked in multiple dimensions of logic integrity, word number requirement and term accuracy, and are automatically corrected according to the result.
7. The automatic bidding document generation method based on large language model and workflow arrangement of claim 1, wherein in step 6, user one-key color rendering is supported, sentence pattern optimization, standard terms and unified style are optimized, and the specialization and logic are improved.

Description

Automatic bidding document generation method based on large language model and workflow arrangement Technical Field The invention relates to the technical field of bidding document generation, in particular to an automatic bidding document generation method based on a large language model and workflow arrangement. Background Currently, in the field of bidding documentation, the main stream technical schemes can be divided into four categories: 1. The traditional manual compiling mode is widely dependent on universal office tools such as Word, excel and the like, and manually collects data, copies historical mark book contents and checks the requirement of the mark-up file one by one. The process generally needs multiple rounds of expert review and manual integration, and has the problems of large workload and frequent version conflict. 2. And the template tool assists in generating, namely, part of software system realizes automatic filling of chapter structures through preset industry templates (such as building and informatization). For example, office plugins such as "golden butterfly bidding assistant", "widely-connected GBC", or SaaS platforms such as "bidding radar", can complete insertion of fixed content by simple rule matching. However, the method relies on a static template library and a weak rule engine, lacks the adaptability to nonstandard bidding requirements, has limited intelligent degree, still requires users to manually import qualification materials and cases, and has insufficient semantic understanding capability. 3. The method adopts a mode of regular rule, TF-IDF keyword extraction, dependency syntactic analysis and the like to perform semi-structural understanding on the bid-bidding document, and generates a bid document framework. However, the method is mainly suitable for the highly standardized industry, and has poor adaptability and low expandability when facing the complex unstructured bidding demands of medical treatment, IT, data centers and the like. 4. General language model (LLM) assisted generation in recent years, users have begun to attempt to directly generate bidding scheme content through prompt word engineering using a general language model such as ChatGPT, meaning thousands, etc. Although the method has a certain natural language organization capability, the problems of large randomness of generated content, poor data consistency (such as inconsistent description of a parameter table and a scheme), fictional facts (such as building qualification certificates), high compliance risk (revealing sensitive information in training corpus) and the like generally exist, and the requirements of real project bidding on professionality, accuracy and traceability are difficult to meet. Although the current bidding documentation work has gradually introduced technical means such as template tools, script automation and general large language models, in general, the following key problems and technical bottlenecks still exist in the existing schemes on the market: 1. The automation degree is low, the process is still highly dependent on manual intervention, and most schemes only realize semi-automatic generation of document fragments at present and lack end-to-end automatic control capability. The user still needs to manually organize the directory structure, call the industry template, import the enterprise qualification information, and check and moisten the bid requirement item by item, so that the whole efficiency is improved only to a limited extent, and the manual operation burden is still heavy. 2. The template has strong dependence and poor flexibility, and template driving tools (such as Word plug-ins or SaaS platforms) commonly used in industry rely on preset chapter structures and fixed formats, so that when facing different fields, different industries or non-standardized bidding projects (such as intelligent buildings, data centers and the like), document structures cannot be dynamically adjusted or targeted technical schemes can be generated, and the suitability and the expandability are poor. 3. The existing scheme generally adopts a single-round text generation mode because of lacking of a context parameter drive and a dynamic response of project elements, and cannot dynamically adjust output contents according to structural characteristics (such as project types, construction units, technical routes, scoring weights and the like) of specific projects, and the capability of 'driving the generation of the contents by the project parameters' is lacking, so that a generated result cannot be attached to a bidding scoring key point, and winning competitiveness is influenced. 4. The semantic understanding capability is weak, the intelligent level is limited, and based on the scheme of an early NLP technology or a static rule engine, the semantic understanding capability is limited to keyword extraction and template filling, and deep logic requirements, soft conditions (