The most common question I get from clients is deceptively simple: should we fine-tune the model, or give it a knowledge base to search?
After deploying both in production this year, the answer is usually: RAG first, fine-tuning second — and only when you know what behaviour you are actually changing.
Retrieval-augmented generation grounds the model in your documents, which fixes the two failure modes that matter most: hallucinated facts and stale answers. It is also cheap to update — edit a document, and the behaviour changes overnight.
Fine-tuning, by contrast, is not a way to teach facts. It is a way to teach behaviour — tone, refusal style, output structure. My rule of thumb: facts live in retrieval, behaviour lives in the weights. Attempting to use one for the other's job is where projects go off the rails.
One more finding from the field: evaluation is the step everyone skips. Every project this year that shipped a test set of 100 edge cases before training finished with a better model than the one with the prettiest loss curve.