The three techniques
| Technique | What it does | Best for |
|---|---|---|
| Prompt engineering | Gives the model clear instructions and examples | General tasks that need no private knowledge |
| RAG | Retrieves relevant passages from your data and gives them to the model | Answers that must come from your documents and stay up to date |
| Fine-tuning | Trains a model further on your examples | A specific tone, format or behaviour that prompts cannot give |
When RAG is the right answer
RAG suits most business questions, because the facts live in your documents and change over time. Updating a document updates the answers, with no retraining.
- Policies, manuals, contracts, product documentation and support history
- Answers that should cite where they came from
- Content that changes often
- Access rules, so people see only what they are allowed to
When fine-tuning is worth it
Fine-tuning teaches a model how to respond, not what is true today. It is useful, and it costs more effort to build and maintain.
- A fixed output format that must be followed reliably
- A particular brand voice or style
- A narrow classification or extraction task with many labelled examples
- Running a small, cheaper model that performs like a larger one on one task
Fine-tuning is a poor way to add facts. Use retrieval for that.
How to decide
- Write 30 to 50 real questions or tasks, with the answers you expect
- Try a prompt first and score it
- If it fails for lack of knowledge, add retrieval and score again
- If it still fails on style or format, consider fine-tuning
- Keep a test set, and rerun it after every change
Whichever you choose, guardrails and evaluation matter more than the technique.
Frequently asked questions
Is RAG better than fine-tuning?
They solve different problems. RAG adds knowledge, fine-tuning shapes behaviour. For answering from your own documents, RAG is usually the better start.
Do we need to train our own model?
Rarely. A hosted or open-source model with prompts and retrieval covers most business needs.
Can we combine them?
Yes. Some systems fine-tune for format and use RAG for facts.
Is our data safe?
We build in your own cloud accounts, with access rules on every source, and use private or open-source models where privacy matters.