When a general-purpose language model isn't good enough for your product, there are two main ways to customise it: retrieval-augmented generation (RAG) and fine-tuning. They solve different problems, and picking the wrong one is one of the most expensive mistakes in AI product development.
The short answer
- Use RAG when the model needs to know facts that are specific to you or change often.
- Use fine-tuning when the model needs to behave in a particular way: a tone, a format or a narrow task it must do reliably.
- Many production systems use both: a lightly fine-tuned model that reads retrieved documents.
Side-by-side comparison
| RAG | Fine-tuning | |
|---|---|---|
| Best for | Knowledge: docs, policies, catalogues | Behaviour: style, format, classification |
| Keeping up to date | Update the documents, instantly live | Retrain the model |
| Citations | Yes, answers link to sources | No, knowledge is baked in |
| Upfront effort | Data pipeline and search tuning | Curated training examples (hundreds+) |
| Risk of made-up answers | Low, when retrieval is good | Higher for factual questions |
| Typical build time | 4–8 weeks | 3–6 weeks plus data preparation |
When RAG is the right call
If your users ask questions whose answers live in documents, whether support articles, product data, contracts or internal wikis, start with RAG. It is cheaper to change, easier to audit and it lets you show exactly where each answer came from. Read our guide to RAG for how it works, or see our RAG development services.
When fine-tuning earns its cost
- Classifying support tickets or emails into your own categories at high volume.
- Writing in a very specific brand voice or structured output format every time.
- Making a smaller, cheaper model perform like a large one on one narrow task.
What about cost?
RAG costs are dominated by engineering the retrieval pipeline and by per-request token usage. Fine-tuning adds data-labelling effort and training runs, and every change means retraining. For most business use cases, RAG delivers value sooner. See how much an LLM app costs for real ranges.
