Well, why not both? If you've already got a tuned model why not use RAG on that to get even better results? It already knows the big picture, it just needs the details so it doesn't have to hallucinate them.