Bidding document automatic generation method based on large language model and workflow arrangement

Abstract

The invention discloses an automatic bidding file generation method based on a large language model and workflow arrangement, which automatically extracts keyword contents such as a'bidding file format 'or a'negotiation response file format' or a'competitive negotiation response file format 'from a bidding file, and identifies a standard directory required by the bidding file. And for the'technical file 'part, automatically constructing a subdirectory according to each scoring point listed in the scoring standard. Scoring detailed rule semantic analysis is introduced, through keyword matching and rule extraction, sub-chapter titles are automatically generated, it is ensured that the technical file content corresponds to scoring points one to one, and the scoring rate is increased. And integrity and accuracy verification is performed on the automatically generated directory structure, and if identification is incomplete or logic is disordered, the system can back the process, re-extract or adjust to ensure that the directory accurately meets the requirements of the bid invitation file.

Description

Technical Field

[0001] This invention relates to the field of tender document generation technology, specifically to an automatic tender document generation method based on a large language model and workflow orchestration. Background Technology

[0002] Currently, in the field of tender document preparation, mainstream technical solutions can be divided into four categories: 1. Traditional manual preparation method: This method relies heavily on common office tools such as Word and Excel. Data is collected manually, historical tender documents are copied, and each requirement is checked against the tender document. This process typically requires multiple rounds of expert review and manual integration, resulting in a large workload and frequent version conflicts.

[0003] 2. Template-based generation: Some software systems automatically fill in chapter structures using preset industry templates (such as those for construction and information technology). For example, Office plugins like "Kingdee Bidding Assistant" and "Glodon GBC," or SaaS platforms like "Bidding Radar," can insert fixed content through simple rule matching. However, these methods rely on static template libraries and weak rule engines, lacking adaptability to non-standard bidding requirements, and have limited intelligence. Users still need to manually import qualification materials and case studies, and their semantic understanding capabilities are insufficient.

[0004] 3. Generation solutions based on early Natural Language Processing (NLP) technologies: These methods employ regular expressions, TF-IDF keyword extraction, and dependency parsing to perform semi-structured understanding of tender documents and generate a tender document framework. However, these methods are mainly suitable for highly standardized industries and have poor adaptability and low scalability when facing complex, unstructured tendering needs in sectors such as healthcare, IT, and data centers.

[0005] 4. General Large Language Model (LLM)-Assisted Generation: In recent years, users have begun to try using general large language models such as ChatGPT and Tongyi Qianwen to directly generate bidding proposal content through prompt word engineering. Although this method has a certain natural language organization capability, it generally suffers from problems such as high randomness of generated content, poor data consistency (e.g., parameter tables do not match the proposal description), fabrication of facts (e.g., fabricated qualification certificates), and high compliance risks (leaking sensitive information in the training corpus), making it difficult to meet the requirements of professionalism, accuracy, and traceability for real project bidding.

[0006] Although current tender document preparation has gradually incorporated technologies such as template tools, script automation, and general-purpose large language models, the existing solutions on the market still generally suffer from the following key issues and technical bottlenecks: 1. Low level of automation: The process still heavily relies on manual intervention. Currently, most solutions only achieve semi-automatic generation of document fragments, lacking end-to-end automated control capabilities. Users still need to manually organize the directory structure, call industry templates, import enterprise qualification information, and proofread and polish each item against the bidding requirements, resulting in limited overall efficiency improvement and a still heavy burden of manual operation.

[0007] 2. Strong template dependence and poor flexibility: Commonly used template-driven tools (such as Word plugins or SaaS platforms) rely on pre-set chapter structures and fixed formats. When faced with different fields, industries, or non-standard bidding projects (such as smart buildings, data centers, etc.), they cannot dynamically adjust the document structure or generate targeted technical solutions, resulting in poor adaptability and scalability.

[0008] 3. Lack of contextual parameter-driven approach, unable to dynamically respond to project elements. Existing solutions typically use a single-round text generation method, which cannot dynamically adjust the output content according to the specific structural characteristics of the project (such as project type, construction unit, technical route, scoring weight, etc.). The lack of "content generation driven by project parameters" results in the generated results failing to align with the key points of the bidding scoring, thus affecting the competitiveness of winning the bid.

[0009] 4. Weak semantic understanding and limited intelligence: Solutions based on early NLP technologies or static rule engines are limited to keyword extraction and template filling, failing to accurately identify deep logical requirements, soft conditions (such as specific judgments of similar performance), or nested constraints (such as conditions for comparing multiple solutions) in tender documents. This is particularly prominent in the generation of unstructured, cross-disciplinary, and multi-chapter linked tender documents.

[0010] 5. The generation based on the Large Language Model (LLM) has the problem of unreliable content. Although the general large language model can achieve preliminary language generation, it lacks accurate data verification and consistency control mechanisms. It often results in output results that do not match the project facts, fictitious data (such as qualifications, awards, team composition) or logical contradictions (such as equipment parameters that are inconsistent with the description). In severe cases, it will lead to bidding compliance risks.

[0011] 6. The inability to establish a unified knowledge management and reuse mechanism: Existing tools often lack enterprise-level "bidding knowledge graphs" or content asset libraries, and lack effective capabilities for reusing historical bids, recommending case studies, and invoking industry standards. This results in the inability to systematically accumulate experience, frequent duplication of effort, and hinders the improvement of long-term bidding capabilities and the establishment of a knowledge loop. In summary, existing solutions either rely heavily on static template generation, depend on weak semantic rule engines, or use unstructured general model calls, none of which can achieve highly reliable, highly consistent, semantically controllable, and data-driven automatic generation of tender documents throughout the entire process. They have obvious shortcomings such as low level of intelligence, weak structural adaptability, and poor industry universality. Summary of the Invention

[0012] This invention addresses the problems of heavy reliance on manual processes, chaotic structure, low content reuse rate, and inability to automatically respond to project elements in the current tender document preparation process. It proposes an automated tender document generation method based on the integration of a large language model and a workflow engine.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: An automatic tender document generation method based on large language models and workflow orchestration is disclosed. The specific steps of this method are as follows: Step 1: Users upload bidding documents. Users upload project bidding documents through the front-end interface. Supported file formats are PDF, Word, and TXT. The system automatically recognizes the document type and performs text parsing. Step 2: Extraction and generation of the directory structure. The system locates the "Tender Document Format" section and extracts the directory structure entries within it; identifies the "Scoring Criteria" section and extracts all clear scoring points, which are then used as the "Technical Documents" subdirectory; and integrates them to generate a complete tender document directory tree, covering the core parts of business response, technical response, and service guarantee. Step 3: Directory verification and repair. Compare with the tender documents and perform logical integrity and structural compliance checks; if the verification fails, return to the previous step to repair the directory. Step 4: Construct prompt words and generate chapter content. For each directory node, construct a structured prompt word template and input the template into the large language model to generate response content. Step 5: Integrate the knowledge base with external search content. Call the enterprise knowledge base to extract content from the enterprise's past bidding documents to fill relevant chapters; call Baidu search nodes to obtain the latest industry specifications, standard terms, or reference texts; and integrate the search results into the prompts in a structured manner through workflow orchestration to improve the rationality and coverage of the generated content. Step 6: Chapter content expansion and polishing. For chapters that do not meet the word count requirement or have thin content, the system can automatically expand the content according to the set threshold.

[0014] Preferably, in step 1, the user can set generation parameters, view the table of contents and chapter content, and make minor adjustments to the content through the front-end interface.

[0015] Preferably, the front-end interface uses a workflow orchestration engine (FastGPT) to control the task flow of each step, and has capabilities such as condition judgment, error rollback, and branch logic control.

[0016] Preferably, in step 4, the structured prompt template includes: chapter title, scoring points, project background, and industry terminology.

[0017] Preferably, in step 5, a large language model (DeepSeek, Tongyi Qianwen large model) is called to generate chapter content, while integrating the enterprise knowledge base and Baidu search results to enrich the semantics of the content.

[0018] Preferably, in step 5, the generated table of contents and chapter content need to be checked in multiple dimensions, including logical integrity, word count requirements, and terminology accuracy, and automatically corrected based on the results.

[0019] Preferably, in step 6, users can polish their work with one click: optimize sentence structure, standardize terminology, unify style, and improve professionalism and logic.

[0020] Compared with the prior art, the beneficial effects of the present invention are: 1. The efficiency of preparation is significantly improved. Compared with the traditional manual method of preparing tender documents, this application uses an automated process to connect the steps of tender analysis, catalog generation, content writing and format output, which greatly reduces the manual operation links.

[0021] 2. Improved professionalism and relevance of generated content: Compared with traditional tools based on fixed templates and rule filling, this application supports semantic parsing of the scoring criteria in the tender documents, automatically extracts key scoring points, and uses this to drive the generation of technical chapter directories and content, ensuring that the generated content accurately corresponds to the scoring elements and enhancing the relevance of the response.

[0022] 3. Enhance the logical consistency and structural integrity of the generated table of contents. Current template-based or AI-based single-round generation tools cannot perform closed-loop verification of the table of contents structure, which is prone to problems such as missing chapters and disordered layout. This application introduces a table of contents logic verification mechanism, which can automatically identify structural missing, repetitive or sequential errors, ensuring the structural coherence of the final generated document and significantly improving the quality of the final delivered document.

[0023] 4. Enhancing Industry Adaptability: By combining a large language model with enterprise-owned knowledge bases and external search engine results, this application constructs a multi-source data fusion mechanism, effectively overcoming the limitations of traditional static templates in terms of content scope and update capabilities. This mechanism can dynamically incorporate professional domain materials, industry terminology, and the latest policy requirements, thereby achieving a simultaneous improvement in the breadth and semantic depth of generated content.

[0024] 5. Supports content reuse and continuous optimization. This system supports the template-based accumulation of directory and chapter structures, supports continuous adjustment and optimization of prompt word strategies, and can gradually accumulate knowledge based on different industry projects to build an enterprise-level bidding response capability library, realizing the transition from "automatic generation" to "intelligent accumulation". Attached Figure Description

[0025] Figure 1 This is a flowchart of the system processing of the present invention; Figure 2 This is an organizational chart of the on-site project department of this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0027] Example 1

[0028] Figure 1 This is a system processing flowchart for the patent. The flowchart illustrates the overall process of an intelligent generation system that automatically processes tender documents based on task type, including three main processing paths: table of contents generation, chapter content generation, and chapter content modification. Details are as follows: First, users upload their bidding documents through the system, which then determines the task type of the uploaded content. If the result is "catalog generation," the system proceeds to the catalog generation branch. The system first identifies the content of the bidding documents and generates a master catalog based on key information such as the "bid document format," "negotiation response format," or "competitive negotiation response document format." Simultaneously, the system extracts the scoring points specified in the bidding documents' evaluation criteria and automatically generates a "technical documents" sub-category. These are then integrated to form a complete bid document catalog. The generated catalog undergoes validation. If validation passes, the catalog is output directly; if validation fails and the current generation count has not exceeded the set limit (e.g., 4 times), the catalog generation process is repeated; if it still fails, the system returns an error message.

[0029] When the task type is "Chapter Content Generation," the system will access content from the knowledge base, process the tender document content, and leverage Baidu search nodes to crawl external materials to enrich the content sources. The system will then determine if the current chapter is a technical document chapter. If it is, it will use the corresponding technical materials to generate the corresponding chapter content; otherwise, it will execute the general chapter generation process. After generating the content, the system will verify the word count of the generated text. If it meets the user's word count requirement, the chapter content will be output; otherwise, the intelligent expansion module will be used to enrich and expand the chapter before outputting it.

[0030] When the task type is "Chapter Content Modification", the system will first load the original chapter content, and then combine the knowledge base content and Baidu search results to generate the modified chapter content, thereby achieving content polishing and semantic enhancement.

[0031] The overall process utilizes multi-path judgment and conditional control to achieve a closed-loop processing mechanism for automatic bid document catalog construction, content generation, and intelligent modification, featuring clear structure, strong controllability, and high adaptability. This system is particularly suitable for application scenarios such as batch bid document preparation, complex project response, and multi-role collaboration.

[0032] To better understand the technical solution of this invention, the following is combined with... Figure 1 The flowchart shown illustrates a detailed description of specific embodiments of the present invention. This embodiment uses water quality monitoring as a basic scenario to describe how, based on user-provided tender documents, a preset process is used to generate the table of contents and content of the tender documents.

[0033] 1. User uploads tender documents Users upload tender documents in PDF or Word format (e.g., "Tender Document: Phase II Project of Online Water Quality Monitoring Equipment and Installation in Guangfeng District.docx"). The system parses the document into structured text and automatically proceeds to the task type identification process.

[0034] 2. Task type determination The system determines the current task type through the interface parameter "generation_type": if it is a "table of contents generation" task, it enters the table of contents processing flow; if it is a "chapter content generation" task, it enters the content generation flow; if it is a "chapter content modification" task, it enters the polishing flow.

[0035] 3. Catalog Generation Process The system reads the chapter format fields from the bidding documents (such as "Tender Document Format", "Negotiation Response Form Format", or "Competitive Negotiation Response Document Format"); based on the title extraction rules described by AI prompts, it identifies the overall directory structure; further, it analyzes the scoring criteria and extracts specific scoring points (such as "After-sales service plan content including ① after-sales service content ② after-sales service plan ③ after-sales service personnel configuration ④ after-sales service personnel training earns 5 points, including any two points earns 2 points, including any one point earns 1 point, and not providing any points earns no points."), and uses this as a subdirectory of the "Technical Documents" section. 8.1 After-sales service plan 8.1.1 After-sales service content 8.1.2 After-sales service plan 8.1.3 After-sales service personnel configuration 8.1.4 After-sales service personnel training Integrate the two directory structures to form a complete tender document directory.

[0036] The directory is validated for completeness using LLM (e.g., whether the "Personnel Organization and Schedule" section is missing); if the validation fails, the system will automatically retry (up to 4 times) and return a prompt message upon failure; if successful, a standardized directory structure will be output for use in the next step of content generation.

[0037] The table of contents generated based on the "Tender Document: Guangfeng District Water Quality Online Monitoring Equipment and Installation Phase II Project.docx" is as follows: 1. Tender documents; 2. Bid Opening Summary Sheet; 3. Itemized quotation sheet; 4. Bid Opening Summary Sheet; 5. Service Request Response / Deviation Table; 6. Business Terms Response / Deviation Table; 7. Qualification and creditworthiness documents that the bidder shall submit; 7.1 A letter of commitment to credit that the bidder shall provide; 7.2 Letter of authorization from the legal representative; 7.3 Bidder's Qualification Statement; 7.4 Consortium Agreement; 7.5 Other qualification documents; 8. Supporting documents required from bidders to implement government procurement policies; 8.1 Declaration Letter for Small and Medium Enterprises; 8.2 Declaration letter from a welfare unit for disabled persons; 9. Technical documents; 9.1 Technical Foundation Response; 9.2 Overall Project Plan; 9.2.1 Overall planning and design; 9.2.1.1 System Architecture; 9.2.1.2 System Functions; 9.2.1.3 Site planning and layout; 9.2.2 Schedule planning; 9.2.3 Installation and debugging plan; 9.3 Technical training program; 9.3.1 Training Objectives; 9.3.2 Training Content; 9.3.3 Training Plan; 9.3.4 Training methods; 9.3.5 Training instructors; 9.3.6 Training effectiveness; 9.4 The proposed person in charge of this project; 9.5 Team members proposed to participate in this project; 9.6 Online water quality analyzer; 9.6.1 Residual chlorine zero-point drift; 9.6.2 Turbidity zero-point drift; 9.7 Ultrasonic flow meter; 9.7.1 Electromagnetic environment level; 9.7.2 Error rate; 9.8 Intelligent miniature circuit breaker; 9.8.1 Resistance strength test; 9.8.2 Uncertainty of AC power; 9.9 Subsequent software integration; 9.10 Basic Business Response; 9.11 After-sales service plan; 9.11.1 After-sales service commitment; 9.11.2 After-sales service plan; 9.11.3 After-sales service guarantee; 9.11.4 Spare parts; 9.11.5 After-sales team; 9.11.6 Response time; 9.12 Bidder's performance record; 10. Other materials; 10.1 Summary of Bidding Companies; 10.2 Materials related to the technical, commercial, and other aspects of the audit score; 4. Chapter Content Generation Process The system accesses the enterprise's internal knowledge base and Baidu search nodes to retrieve relevant background materials based on the title or user-specified content. It then determines whether the chapter title provided by the user is a "technical document" chapter: if so, it invokes the "technical document content generation" specialized model; otherwise, it uses a general text generator. After generating the text, it performs a word count check; if the word count does not meet the requirements, it uses an intelligent expansion function to complete the text; finally, it outputs a structured chapter draft for the user to review or further refine.

[0038] For example, the content of "Chapter: 9. Technical Documents - 9.2 Project Overall Plan - 9.2.1 Overall Planning and Design - 9.2.1.1 System Architecture" is generated as follows: The system architecture design must fully consider the overall project requirements and the feasibility of technical implementation to ensure the system's efficiency, stability, and scalability. The following are the main components of the system architecture design: 1. Infrastructure layer Cloud platforms: Adopting public cloud, private cloud, or hybrid cloud deployment models, such as Alibaba Cloud, Tencent Cloud, and Huawei Cloud, to provide elastic computing, storage, and network resources.

[0039] Containerization: Using Docker container technology, combined with Kubernetes (K8s) for container orchestration, to achieve rapid deployment, scaling and management of services.

[0040] Microservice architecture: Microservices are built based on frameworks such as Spring Cloud and Dubbo to achieve service decoupling, independent deployment, and scalability.

[0041] 2. Data Acquisition and Transmission Message queues: Utilize message middleware such as Kafka and RabbitMQ to achieve asynchronous data transmission and decoupling.

[0042] ETL tools: Use tools such as Apache Nifi and DataX to automate the data extraction, transformation, and loading process.

[0043] Log collection: Use tools such as Flume and Logstash to collect semi-structured data such as system logs and application logs.

[0044] 3. Data storage and processing Data warehouse: Hadoop HDFS is used to store large-scale data, while Hive serves as a data warehouse, providing SQL query capabilities.

[0045] Real-time computing: Apache Flink is used for real-time data stream processing, supporting complex event processing and window computing.

[0046] Batch processing: Apache Spark provides powerful data processing capabilities, supporting advanced analytics such as machine learning and graph computing.

[0047] NoSQL databases, such as MongoDB and Cassandra, are used to store unstructured or semi-structured data and support high-speed read and write operations.

[0048] 4. Data Management and Governance Data governance measures include data collection, cleaning, storage, backup, and recovery mechanisms to ensure data quality and security.

[0049] Security measures: Strengthen data encryption, access control, audit trails and other measures to protect personal information and sensitive data, and comply with national information security laws and regulations.

[0050] The above architecture design ensures that the system can meet current and future business needs, while also possessing good scalability and maintainability.

[0051] 5. Chapter Content Modification Process The system loads the historical chapter content entered by the user; at the same time, it calls the knowledge base and Baidu search interface to extract contextual information related to the original text; it uses a language model to perform logical error correction, semantic enhancement and format optimization; and generates a polished version to return to the front end.

[0052] For example, the user inputs: Paragraphs to be optimized: To ensure the smooth implementation of the project, while achieving standardized and regulated management, improving project execution efficiency, standardizing the project implementation process, and ensuring that the undertaken project is completed with high quality, an on-site project team was established and equipped with experienced personnel.

[0053] User requirements: Add images to enrich the content Output after LLM editing and polishing: To ensure the smooth implementation of the project, achieve standardized and regulated management, improve project execution efficiency, and guarantee high-quality completion of the undertaken project, our company has decided to establish an on-site project team, staffed with experienced professionals. The on-site project team will be fully responsible for the planning, organization, coordination, and control of the project, ensuring its orderly progress according to plan.

[0054] In addition, to more intuitively showcase project progress, team collaboration, and interim results, images will be incorporated into relevant documents, reports, and presentation materials. These images may include on-site construction photos, team member work scenes, and visual representations of key achievements. The visual presentation of images further enriches the content and enhances the effectiveness of information delivery. An organizational chart of the on-site project team, for example... Figure 2 .

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatically generating tender documents based on a large language model and workflow orchestration, characterized in that: The specific steps of this method for automatically generating tender documents based on large language models and workflow orchestration are as follows: Step 1: Users upload bidding documents. Users upload project bidding documents through the front-end interface. Supported file formats include PDF, Word, and TXT. The system automatically recognizes the document type and performs text parsing. Step 2: Extraction and generation of directory structure. The system locates the "Tender Document Format" section and extracts its directory structure entries; identifies the "Scoring Criteria" section and extracts all clear scoring points, which are then used as the "Technical Documents" subdirectory; and integrates them to generate a complete tender document directory tree, covering the core parts of business response, technical response, and service guarantee. Step 3: Directory verification and repair. Compare with the tender documents and perform logical integrity and structural compliance checks; if the verification fails, return to the previous step to repair the directory. Step 4: Construct prompt words and generate chapter content. For each directory node, construct a structured prompt word template and input the template into the large language model to generate response content. Step 5: Integrate the knowledge base with external search content, call the enterprise knowledge base, and extract the content of the enterprise's past bidding documents to fill in the relevant chapters; Call Baidu search nodes to obtain the latest industry standards, standard terminology, or reference texts; By orchestrating workflows, search results are structured and integrated into prompts, improving the rationality and coverage of generated content. Step 6: Chapter content expansion and polishing. For chapters that do not meet the word count requirement or have thin content, the system can automatically expand the content according to the set threshold.

2. The method for automatically generating tender documents based on a large language model and workflow orchestration as described in claim 1, characterized in that: In step 1, users can set generation parameters, view the table of contents and chapter content, and make minor adjustments to the content through the front-end interface.

3. The method for automatically generating tender documents based on a large language model and workflow orchestration as described in claim 1, characterized in that: The front-end interface uses a workflow orchestration engine to control the task flow of each step, and has the ability to perform conditional judgments, error rollback, and branch logic control.

4. The method for automatically generating tender documents based on a large language model and workflow orchestration as described in claim 1, characterized in that: In step 4, the structured prompt template includes: chapter title, scoring points, project background, and industry terminology.

5. The method for automatically generating tender documents based on a large language model and workflow orchestration as described in claim 1, characterized in that: In step 5, the large language model is invoked to generate chapter content, while integrating the enterprise knowledge base and Baidu search results to enrich the semantics of the content.

6. The method for automatically generating tender documents based on a large language model and workflow orchestration as described in claim 1, characterized in that: In step 5, the generated table of contents and chapter content need to be checked from multiple dimensions, including logical integrity, word count requirements, and terminology accuracy, and then automatically corrected based on the results.

7. The method for automatically generating tender documents based on a large language model and workflow orchestration as described in claim 1, characterized in that: In step 6, users can polish their work with one click: optimize sentence structure, standardize terminology, unify style, and improve professionalism and logic.

Images