Is it a very large codebase? Anything else distinctive about it? Are you often asking high-level/conceptual questions? Those are the questions that would help me understand why you might be seeing better results with RAG.