CN-122018965-A - Large-model-based embedded software document generation intelligent agent construction method

CN122018965ACN 122018965 ACN122018965 ACN 122018965ACN-122018965-A

Abstract

The invention discloses an intelligent agent building method for generating embedded software documents based on a large model, which comprises the specific processes of knowledge structuring and prompt word design, decomposing a general requirement of software development documents into a rule set which can be understood and guided by a machine and designing a prompt word template, building an intelligent agent framework, building a document generating process by taking task planning and execution as cores, integrating all modules of the built intelligent agent framework, and performing closed-loop verification. According to the document generation agent construction method, through intelligent guidance, content generation and quality inspection, the efficiency, accuracy and standardization of writing a software document by a software designer according to the standard of the general requirement of software development document are obviously improved, and the requirement analysis capability is enhanced.

Inventors

ZHANG XIAOQIAN
LIU HUI
LUO HAO
LIU JIAXIANG

Assignees

湖北三江航天红峰控制有限公司

Dates

Publication Date: 20260512
Application Date: 20251217

Claims (6)

1. The method for constructing the intelligent agent generated by the embedded software document based on the large model is characterized by comprising the following specific steps of: s1, knowledge structuring and prompt word design, namely decomposing a general requirement of a software development document into a rule set which can be understood and guided by a machine and designing a prompt word template; S2, constructing an agent framework, and constructing a document generation flow by taking task planning and execution as cores; And S3, integrating all modules of the intelligent agent framework built in the step S2, and performing closed-loop verification.
2. The method for constructing the large model based embedded software document generating agent of claim 1, wherein the specific flow of step S1 is as follows: S11, deeply deconstructing the standard, and converting standard clauses into a specific and operable rule and problem list; S12, constructing a prompt word template library, and designing basic prompt word templates aiming at different scenes including document generation, question and answer and quality inspection; S13, carefully selecting an example sample, and collecting and writing partial high-quality document fragments to serve as examples in sample prompts.
3. The method for constructing the large model based embedded software document generating agent of claim 1, wherein the specific flow of step S2 is as follows: s21, selecting and designing a framework, adopting an agent framework, and designing a task planner, a tool set and a memory module based on the selected agent framework; S22, an intelligent interaction guide module is constructed, multi-round dialogue logic is realized, and the problem to be queried or the action to be executed in the next step is dynamically determined according to the current input and the historical context of a user; s23, constructing a content generation and completion module, integrating the prompt word template constructed in the step S1 into an intelligent agent, and realizing the functions of 'one-key generation chapter' and 'one-key quality initial check'; S24, constructing a quality inspection and consistency verification module, and enabling the large model to carry out compliance judgment, term correction and logic consistency inspection on the existing text based on an inspection list through sample classification and text comparison prompt technology; S25, a dynamic context management module is constructed and used for storing and managing project contexts, and consistency of cross-section information is achieved.
4. The method for building an agent for generating an embedded software document based on a large model as recited in claim 3, wherein in step 321, any one of LANGCHAIN, LLAMAINDEX is selected as said agent framework.
5. The method for building an agent for generating a large model based embedded software document as recited in claim 3, wherein in step 321, said tool set comprises a hint word template executor and a vector database retriever.
6. The method for constructing the large model based embedded software document generating agent of claim 1, wherein the specific flow of step S3 is as follows: S31, developing a user interface, providing a concise man-machine interaction interface, and facilitating interaction between a user and an intelligent body; s32, performing end-to-end test verification, namely selecting a past typical project, generating a complete set of documents by using an intelligent agent from scratch, and comparing the complete set of documents with manually written documents; And S33, iterating and optimizing the prompt word, and continuously adjusting and optimizing the prompt word template according to the test verification result.

Description

Large-model-based embedded software document generation intelligent agent construction method Technical Field The invention relates to the technical field of embedded document generation, in particular to a large-model-based embedded software document generation intelligent body construction method. Background As embedded software plays an increasingly critical role in modern construction, its quality, reliability and maintainability become critical. The file of the embedded software is used as a blueprint and a description of the software development process, and is a core basis for ensuring the quality of the software and facilitating subsequent testing, maintenance and upgrading. Various software documents are compiled in the industry by strictly referring to the standard of general requirement of software development documents. However, in the practical process, serious challenges are faced in that software designers have larger difference in understanding depth of standards and experience of document writing, so that document quality is unstable, implicit requirements, boundary conditions or abnormal handling are easy to miss when documents such as software requirement specification are written, hidden dangers are buried for subsequent development, a large number of errors which can be avoided, such as inconsistent formats, non-uniform terms, section omission, description ambiguity and the like, exist in the document writing process, the specialization and readability of the documents are greatly reduced, the purposes of writing each document in the standards are greatly reduced, the understanding of content depth and mutual relevance is insufficient, the documents often flow in a form, and the engineering value of the documents cannot be really exerted. The problems seriously restrict the improvement of the research and development efficiency of the embedded software and the guarantee of the product quality. Disclosure of Invention Aiming at the technical problems, the invention provides a large-model-based embedded software document generation intelligent agent construction method, which remarkably improves the efficiency, accuracy and standardization of writing a software document by software designers according to the standard of general requirement of software development document and strengthens the requirement analysis capability through intelligent guidance, content generation and quality inspection. The method for constructing the intelligent agent generated by the embedded software document based on the large model comprises the following specific processes: s1, knowledge structuring and prompt word design, namely decomposing a general requirement of a software development document into a rule set which can be understood and guided by a machine and designing a prompt word template; S2, constructing an agent framework, and constructing a document generation flow by taking task planning and execution as cores; And S3, integrating all modules of the intelligent agent framework built in the step S2, and performing closed-loop verification. As a preferable mode of the above technical solution, the specific flow of step S1 is as follows: S11, deeply deconstructing the standard, and converting standard clauses into a specific and operable rule and problem list; S12, constructing a prompt word template library, and designing basic prompt word templates aiming at different scenes including document generation, question and answer and quality inspection; S13, carefully selecting an example sample, and collecting and writing partial high-quality document fragments to serve as examples in sample prompts. As a preferable mode of the above technical solution, the specific flow of step S2 is as follows: s21, selecting and designing a framework, adopting an agent framework, and designing a task planner, a tool set and a memory module based on the selected agent framework; S22, an intelligent interaction guide module is constructed, multi-round dialogue logic is realized, and the problem to be queried or the action to be executed in the next step is dynamically determined according to the current input and the historical context of a user; s23, constructing a content generation and completion module, integrating the prompt word template constructed in the step S1 into an intelligent agent, and realizing the functions of 'one-key generation chapter' and 'one-key quality initial check'; S24, constructing a quality inspection and consistency verification module, and enabling the large model to carry out compliance judgment, term correction and logic consistency inspection on the existing text based on an inspection list through sample classification and text comparison prompt technology; S25, a dynamic context management module is constructed and used for storing and managing project contexts, and consistency of cross-section information is achieved. As a preferred embodiment of the foregoing disclosure, in step