Evaluating RAG for large scale codebases