Code Auto-generation System based on Large Language Models

This paper introduces a large-scale model-based code automatic generation system and its construction method. It aims to use deep learning and natural language processing techniques, based on pre trained large language models, to automatically understand user input and needs, and generate executable program code. In addition to supporting automatic generation of general code, this patent is more suitable for professional system development scenarios, especially those with specific requirements for generating code. For example, in the front-end development process, users need to use previously created custom controls, custom styles, or layouts in the automatically generated code; Or when accessing interface services on the backend, users need to automatically generate code for interfaces with specific logical requirements, such as specific parameter concatenation methods and encryption methods. The core of the system in the paper is a dedicated code generator. When users describe their needs, they provide a simple code example. The system decomposes the generation process into multiple steps based on requirements and examples, analyzes the dependencies between these steps, generates prompts, and finally solidifies the entire generation process into a dedicated code generator. Different scenarios correspond to different specialized code generators. This system does not rely on specific large-scale language models, nor does it require special training for the code. It is convenient and fast, can significantly shorten development time, flexibly handle changes in requirements in various specific scenarios, reduce repetitive work, and standardize system development.