Fine-tuning

Fine-tuning means taking an existing model and training it further on your own examples, so the behavior you want gets baked in. Instead of spelling out the format and tone in a long prompt every time, you show it a few hundred or few thousand examples of the input and the ideal output, and it learns the pattern. The result is a model that does your specific thing more reliably, often faster and cheaper per call, since you no longer need giant prompts.

It's more effort than prompting. You need a clean dataset of good examples, and you have to retrain when things change. So it's a step you take once prompting and RAG have carried you as far as they can.

Fine-tune when you know exactly what you want the model to do and can show it in examples, and prompting alone keeps falling short: a consistent format it won't hold, or a narrow task you run at high volume where a shorter prompt saves real money.

Skip it when what you need is knowledge rather than behavior, since RAG covers that. And hold off while your requirements are still moving, since you'd be retraining constantly. The usual order: push prompting as far as it goes, add RAG when you need facts, and fine-tune only when both leave you short.