Retrieval-Augmented Generation for Scientific Code Understanding

Small local models can answer questions about a scientific codebase when repository knowledge is prepared in advance. This study parses code, maps relationships, and generates explanations before storing that material in a searchable index. At question time, the assistant retrieves relevant material for its answer.

On 100 questions about the IPPL C++ codebase, an automated judge rated the best tested model's answers as 70 fully correct and 19 partially correct. The system explains code rather than editing or testing it. Its useful tradeoff is upfront preparation for repeated local question-answering, with the index needing maintenance when the underlying code changes.