Stop asking for fine-tuning. You almost certainly need retrieval.

At least once a week, a member asks how to fine-tune a model on their company data. When we ask what problem they're solving, it's almost always the same: 'the model doesn't know our products / policies / pricing.' That's a knowledge problem, and fine-tuning is the wrong tool for knowledge problems.

Why fine-tuning fails at knowledge

Fine-tuning adjusts how a model behaves — its style, format, and task-specific patterns. It's poor at reliably storing facts, and worse, the facts it does absorb are frozen at training time. Your pricing changes next quarter; your fine-tuned model doesn't. You're now maintaining a model that confidently recites outdated information — the worst failure mode in business AI.

Retrieval-augmented generation keeps knowledge outside the model, where it can be updated by editing a document. The model looks up current information at question time. When the policy changes, you change the document, and every answer after that reflects it.

The ladder

Work through these in order, and stop at the first rung that solves your problem: better prompting (most 'model problems' are instruction problems), few-shot examples (style and format), RAG (knowledge), model routing (cost), and only then fine-tuning — for high-volume, narrow tasks with abundant labeled data where the cheaper rungs demonstrably fell short.

In three years of member projects, fewer than one in twenty teams that thought they needed fine-tuning actually did. The ones that did had thousands of labeled examples and a narrow, stable task. If that's not you, climb the cheaper rungs first.