Fine-tuning takes a pretrained model and trains it further on your own examples, updating the weights so the new behavior becomes part of the model. It is how you teach a model a consistent format, tone, or task shape that prompting alone struggles to enforce.
Why it matters in production
Fine-tuning is good at behavior and bad at knowledge. If you need every answer in a specific structure, or a smaller cheaper model to mimic a larger one's outputs on a narrow task, fine-tuning earns its keep. What it does not reliably do is teach the model facts — and confusing those two is the expensive mistake.
The common mistake
Fine-tuning to fix "the model doesn't know about our product." That is a retrieval problem, not a training one. Bake facts into weights and they go stale the moment your docs change, and you cannot cite a source. RAG updates by re-indexing a file; fine-tuning updates by running a training job and re-validating the whole model. For knowledge, retrieval wins almost every time.
How we actually use it
We reach for fine-tuning to lock in format and behavior, not to inject facts — and only after prompting and retrieval have hit a real ceiling, because it adds a training pipeline and a versioned artifact to maintain. When we do fine-tune, evals are non-negotiable: a fine-tune can silently regress on cases outside its training set, and only a test set catches it. Changing the base model later is a full model migration, fine-tune included.