Fine-tuning

Fine-tuning means taking an already-trained AI model and continuing its training on a smaller set of examples from your own domain: input/output pairs that show exactly how you want it to respond. The model keeps everything it already knew and absorbs the behavior in your examples: tone of voice, format, vocabulary, classification decisions.

It differs from RAG: RAG gives the model knowledge at question time (retrieves documents and injects them into the prompt); fine-tuning changes the model's own behavior. Rule of thumb: information that changes often → RAG; style, format and well-defined repetitive tasks → fine-tuning. Providers like OpenAI offer fine-tuning as an API service; open models can be tuned on your own or rented GPU infrastructure (Replicate, RunPod).

The costs: preparing hundreds or thousands of quality examples, training and evaluating, which is why it pays to start with prompt engineering and RAG, and only move to fine-tuning when they fall short. A concrete example: a fintech tunes a model on thousands of customer messages already labeled by category; the fine-tuned model then classifies new tickets in the company's exact standard, with short prompts and lower cost per call.