Fine-tuning updates a model's weights with additional training data. It is useful for changing a style, a format, a classification boundary, or a tool-calling habit that prompting cannot stabilize. It is a poor way to add facts, because a model trained to reproduce an answer may reproduce it confidently when it is wrong, and the knowledge goes stale the moment the source changes.
The practical variants differ in cost. Full fine-tuning updates every weight and needs the memory to match; parameter-efficient methods such as low-rank adapters train a small number of extra parameters and are cheaper to run and store. Either way, you need a training set, a held-out set, and an evaluation that measures the behavior you wanted to change, not just the loss.
Fine-tuning is not a substitute for retrieval when the requirement is "answer from these documents," and it is not a substitute for a skill when the requirement is "follow this procedure." It changes tendencies, not guarantees. See RAG and test a model on your own tasks.