Compare · Building · Research notes
RAG vs fine-tuning — scored
Fine-tuning is the heavier lever. Start with prompts, add RAG for changing facts, fine-tune for stable style or format.
Last reviewed: 2026-08-15. Re-run the protocol after major model or pricing changes.
How do you choose RAG vs fine-tuning?
- Write whether the facts change weekly.
- Write whether you need a stable voice/format.
- Price retrieval ops vs a training run and eval set.
- Keep a holdout. Do not trust a single demo.
What is the scoring rubric?
| Criterion | Weight | How to score |
|---|---|---|
| Facts change | Gate | If yes, RAG (or search) before fine-tune. |
| Style/format lock | 1–5 | Fine-tune or a strong prompt pack. |
| Eval set exists | Gate | Fail fine-tunes without a holdout. |
| Cost to update | 1–5 | Re-index vs re-train. |
Which lever should I pull?
| Tool | Best fit | Use when | Avoid when | Note |
|---|---|---|---|---|
| Prompting | First lever | The job fits in context | Facts are long or changing | Always start here. |
| RAG | Changing documents | You can retrieve and cite | You only needed a tone change | Evaluate groundedness. |
| Fine-tuning | Stable style/format | You have a clean train/holdout set | Facts go stale weekly | Combine with RAG when both apply. |
Learn the craft, not just the tool
Practice the transferable habit in Fine-tuning vs RAG, then Three-way decision, Fine-tune guide.
FAQ
Should I fine-tune or use RAG?
RAG for changing facts. Fine-tune for stable style or format. Many products need both.
Can prompts replace both?
Often for a first version. They fail when the corpus is large or the format must be locked.