Fine-tuning does not give the model knowledge, it gives it behaviour. If you want it to work in your format, in your voice, with your classification logic, this is the right tool. If you want to put new facts into it — company documents, current prices — then this is not what you need; RAG is.
Where it works, it works spectacularly. A fine-tuned model with 350 million parameters reached a 77.6% success rate on a tool-calling task, while a general model with 175 billion parameters managed 26% on the same task. Five hundred times fewer parameters, three times the result — that is the real promise of fine-tuning.
The most common misconception is that it is free. It is not: forgetting grows with model size, and in one experiment the fine-tuned model degraded on completely unrelated questions too. On top of that comes provider risk — one large provider is retiring its fine-tuning platform right now and no longer takes new users.
When it is worth it and when it is not is worked through with numbers in the fine-tuning trade-off, a separate article. The directly related terms: RAG, LoRA and instruction tuning.