Fine-tuning is training an existing AI model further on your own examples so it specializes in a task. It changes the model's weights, so the change stays in the model: it learns to follow the tone, format or judgment of those examples without you repeating them in every prompt.
A question from the app. Pick an answer.
A store wants its bot to quote prices that change every week. What works best?
It can learn some prices, but good recall is not guaranteed. And every change means retraining.
The data is right in front of it on every answer. A price changes, you change the file, and the model stays the same.
Being bigger or newer does not give it your current prices. Without checking them, it may get them wrong or make them up.
You start from a model that already learned the language and a lot about the world. You prepare examples for it: a request and the ideal answer, over and over. Training on them nudges its weights toward that way of answering. The result is a new version of the model, which you use just like the original.
Many LLMs go through fine-tuning before release: that is how they learn to follow instructions and reply like an assistant. Techniques like LoRA freeze the original weights and train small matrices added to the model, which costs less.
A store wants a bot. Be friendly and keep it short: that goes in the prompt, and you can test it in minutes. Know prices and stock, which change every week: that is RAG, the bot looks up the current list before answering. Keep the brand's tone and the format the order system reads more consistent, when the prompt falls short: that is where fine-tuning can help.
The short rule: fine-tuning changes how it answers; RAG gives it data it does not have. They can be combined, and it almost always pays to try the prompt first.
Weights are not a database. If you train on your price list, the model can learn facts, but it is not guaranteed to recall them well and it does not update itself: every change means retraining. For prices and stock, have the bot check the current source. With RAG it searches your documents by meaning and, if a price changes, you update the file and its index.
Costs include preparing and checking good examples, training and running the tuned model. Mistakes in the examples can carry over to the model. Also, not every model can be fine-tuned, and on some platforms the tuned model costs more per token than the original. Before deciding, work out what it costs against a good prompt.