When retrieval is the right answer
RAG, long context, fine-tuning and tools solve different problems.
RAG injects relevant external knowledge at query time. Use it when the corpus is too large for the window, changes frequently, needs per-user access control, or requires citations. Long context — just putting the documents in the prompt — is simpler and often better when the corpus is small and stable; do not build a vector database for 40 pages of policy. Fine-tuning changes style, format and task behaviour; it is a poor and expensive way to inject facts. Tools/APIs are correct for live, precise, structured data — never embed your orders table, query it.
'The model doesn't know our current inventory' is not a RAG problem. It is a tool-use problem. Fresh, exact, structured data comes from an API call.
- Small stable corpus → long context (plus caching). Large or changing corpus → RAG.
- Fine-tuning teaches behaviour, not facts.
- Live structured data belongs behind a tool call, not an embedding index.