Enhancing the ability of LLMs for spaceborne equipment code generation via retrieval-augmented generation and contrastive learning

In the code generation field, Large Language Models (LLMs) pre-trained on numerous open-source code fragments show powerful reasoning abilities and remarkable downstream performance. They assist code generation by combining retrieval techniques like retrieving relevant code fragments as templates or using retrieval results to supplement natural language descriptions and get code examples. However, in domains like aerospace equipment, existing code generation technologies perform suboptimally. Different aerospace equipment has different functions and significant data processing and loading differences. There is a lack of effective retrieval methods to provide semantically similar code contexts for LLMs, hindering code generation from meeting complex task requirements. To address this, we propose CodeCLARE, a retrieval-augmented code generation framework. It first fine-tunes UniXcoder via contrastive learning and uses it as a semantic encoder for code fragment retrieval. Then, the NL2Code search strategy is adopted with program requirements as queries. In the final stage of the code generation process, through a “Few-Shots Selection” mechanism, the prompt templates effectively integrate both the retrieved code examples and the specific requirement information, enabling the successful generation of highly accurate C++ code through the advanced capabilities of LLMs. Experimental results show that our approach significantly improves code quality compared to traditional ones and provides an effective solution for spacecraft control code generation.