Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories
Documentation built around code dependencies improved text-to-code retrieval in tests on OpenCV and a proprietary 5-million-line repository.
Code Comprehensor combines a dependency call graph, LLM-generated hierarchical documentation and a model trained to align text with code. It retrieves relevant code and surrounding context as evidence for an LLM answering domain-specific questions. Adding parent and sibling context, hierarchical documentation and synthetic questions consistently improved retrieval across both repositories. The authors report faithfulness scores above 90% from an LLM-as-judge evaluation using RAGAS.
ArXiv · AI/CL/LG's note
Code Comprehensor combines a dependency call graph, LLM-generated hierarchical documentation and a model trained to align text with code. It retrieves relevant code and surrounding context as evidence for an LLM answering domain-specific questions. Adding parent and sibling context, hierarchical documentation and synthetic questions consistently improved retrieval across both repositories. The authors report faithfulness scores above 90% from an LLM-as-judge evaluation using RAGAS.
ArXiv · AI/CL/LG's note
score 4