If you’re “searing” specific documents and incidents, would it make more sense to use either vector search or a retrieval (RAG) model to pull from these sources, instead of generating new text? Even if you have a lot of data it may be difficult to fine-tune LLaMa to generate reliable responses.