Prompting, RAG or fine-tuning: which one do you need?
People often ask for a fine-tuned model when they really need better instructions, or a search step. The three approaches solve different problems, and the cheapest one often does the job. Here’s how we think about the order.
Try them in this order
Figure 1 · A sensible sequence
Write a clear promptwith examplesAdd retrievalif it needs your documentsConsider fine-tuningif behaviour stays inconsistent
What each one is for
Figure 2 · Choosing an approach
| Approach | Good for | Weak at | Effort |
|---|---|---|---|
| Prompting | Tone, format, task instructions, examples | Facts the model doesn’t know | Low |
| RAG | Answers from your own, changing documents | Changing how the model writes or reasons | Medium |
| Fine-tuning | A consistent style or format at scale | Teaching new facts reliably | High |
The common mix-ups
- Fine-tuning to add knowledge. It shapes behaviour better than it stores facts. For facts that change, retrieval is the more reliable route, as in our RAG explainer.
- RAG to fix style. Retrieval won’t change how the model writes. A better prompt will.
- Skipping measurement. Without a test set, you can’t tell whether a more complex approach helped. Start with twenty real examples.
A quick way to decide
Ask what is actually failing. If the answers are in the wrong format, fix the prompt. If they’re wrong about your own data, add retrieval. If they’re right but inconsistent in style or structure across thousands of calls, then look at fine-tuning. Most projects never need the last step.