Fine-tuning takes a model that already knows language in general and continues training it on a smaller, task-specific dataset so it specialises — in a domain’s vocabulary, a particular format, or a consistent tone. It changes the model’s weights, unlike retrieval, which changes only what you show the model at query time.
It is the right tool when you need consistent behaviour that prompting cannot reliably produce, and the wrong one when your real problem is access to current facts (that is retrieval) or when you have too little quality data to train on. Deciding between fine-tuning and retrieval — often the answer is retrieval — is one of the first real architecture calls in an AI project.