I remember seeing somewhere that LLMs performed on a niche subset of data perform worse than general ones. RAG is probably best here.