akira

CodeRAG-Bench: Can Retrieval Augment Code Generation?

NAACL Findings · 2025

summarized 2026-05-15

Summary

RAG was observed to increase performance of LMs on text-oriented tasks, but its capabilities for coding specifically are underexplored. CodeRAG-Bench introduces a benchmarking dataset for evaluating LMs and RAG on coding tasks specifically. The paper finds that retrieving high-quality code improves code generation, but retrievers often struggle to fetch useful contexts, and generators sometimes struggle to use that context effectively.

Notes

The context is appended to the prompt.

Questions