What's the best way to use an LLM with a large codebase that isn't RAG? Ideally we could have the full source in the context or already trained into the model... I was thinking I could set something to fine tune a model overnight and every morning I'd have a fresh one ready. Any ideas?