RAG vs Fine-Tune — A Decision Frame

The "should I fine-tune?" question is the wrong question 90% of the time.

Default to RAG

Retrieval-augmented generation is faster to set up, cheaper to maintain, easier to update, and good enough for most use cases.

When to fine-tune

• You need consistent output style, not facts.
• You're operating at scale (millions of calls) and prompt size is the bottleneck.
• You need to teach a new task that prompting can't reliably produce.

The false third option

Founders often pick "fine-tune on our docs" as a way to feel productive. This is RAG with extra steps and worse results. Don't.

A 3-step audit

1. Can a sharp prompt get 80% of what you want? Stop there.
2. Does the gap come from missing knowledge? → RAG.
3. Does the gap come from missing skill/style? → fine-tune.