En bref
Fine-tuning is the process of taking a pretrained AI model and training it further on a focused set of your own examples, so it reliably adopts a specific style, format, or skill without being told every time in the prompt. The base model already knows language broadly; fine-tuning nudges its weights toward your narrow use case, whether that is answering in your brand voice, always returning a certain JSON shape, or handling a specialised domain. It is a heavier tool than prompting or RAG because it changes the model itself, so it pays off when you need consistent behaviour at scale and you have good example data to teach from.
How fine-tuning works
Fine-tuning trains an existing model on labelled examples of the behaviour you want. You prepare pairs of input and ideal output, run a training job, and get back a customised version of the model.
- Collect examples: gather high-quality input-to-output pairs that show the behaviour you want.
- Train: run a fine-tuning job that adjusts the model weights toward those examples.
- Use: call your customised model, which now leans toward that style or task by default.
Fine-tuning vs RAG vs prompting
These three ways to adapt a model solve different problems. Prompting shapes one response with instructions. RAG adds fresh knowledge at query time by retrieving documents. Fine-tuning bakes in behaviour by changing the weights. A simple rule: prompt first, add RAG when the model needs to know your facts, and fine-tune when you need consistent style or format that prompting cannot hold. Knowledge belongs in RAG; behaviour belongs in fine-tuning.
When fine-tuning is worth it
Fine-tuning is worth the effort when a task repeats at scale, the desired output is narrow and consistent, and you can supply enough clean examples to teach it. It is usually not the first move: it costs time and money, it must be redone when the base model updates, and it does not add live knowledge. For most builders, a sharp system prompt plus RAG gets you most of the way, and fine-tuning is the last step for the cases where consistency really matters.