Glossary

Fine-tuning

Fine-tuning is the process of further training an existing language model on your own examples so it adopts a specific style, format, or skill. It changes the model's weights, baking the behavior in, rather than supplying knowledge at query time the way RAG does.

You start from a pretrained model and train it on hundreds or thousands of input-output pairs that show exactly how it should behave — a tone of voice, a strict output format, a classification it keeps getting wrong. The result is a model that does that one thing more reliably and often with shorter prompts. What fine-tuning is not good at is keeping facts current: anything baked into weights is frozen until you retrain.

When building a product, fine-tuning is rarely the first move and often not needed at all. A well-written prompt plus RAG solves most problems faster and stays editable. Reach for fine-tuning when you have a high-volume, narrow task where prompt instructions aren't enough, or when you need consistent structured output at scale. The cost is real: you need a clean labelled dataset, evals to prove it actually improved, and a plan to retrain as your needs shift. We treat it as a targeted tool, not a default.

// faq

Frequently asked questions

Do I need to fine-tune to get good results?
Usually not. Most teams get further, faster, with a sharp prompt and RAG, both of which stay editable and cite sources. Fine-tuning earns its keep on narrow, high-volume tasks where prompting alone can't hold a format or style consistently. Try prompting and retrieval first, measure with evals, and only fine-tune when the data shows a clear gap.
Start a project inquiry