When should you fine-tune instead of using RAG?

**Fine-tuning** and **RAG** look similar from the outside (both make the model better at your task) but they fix different things. Fine-tuning updates the model's weights on your examples. After fine-tuning, the model itself behaves differently, it adopts a tone, follows a format, reasons in a domain specific way, or learns a new skill like JSON output. You're permanently changing the brain. RAG keeps the model frozen and instead retrieves relevant text from a vector database at query time, then stuffs that text into the prompt. The model now answers grounded in those documents. Nothing about the model has changed, the knowledge lives in the index. A useful rule of thumb: if the issue is HOW the model talks or reasons, fine-tune. If the issue is WHAT facts it knows, use RAG. Tone, format, refusal behavior, and domain reasoning go in fine-tuning. Product docs, policy text, customer records, and yesterday's news go in RAG. They are not exclusive. A production chatbot might be fine-tuned for company voice and refusal policy, then use RAG to answer questions from a constantly updating help center.

**Fine-tuning** changes what the model IS; **RAG** changes what the model CAN SEE at query time. They solve different problems. Fine-tune when you need to shift model behavior. Examples: a custom tone, structured output format, domain specific reasoning patterns, refusal rules. A skill that requires gradient updates fits here too (function calling style, JSON schema adherence, multi-step domain logic). LoRA or QLoRA makes this cheap on top of an open weight base. Use RAG when the problem is missing knowledge. Think private data the model never saw. Use it when knowledge changes faster than you retrain (prices, inventory, news), or when corpora are too large to fit any context window. Updating the index is minutes; retraining a model is days. The trade off: RAG adds latency and tokens per query but stays fresh. Fine-tuning adds training cost but ships cheaper, lower latency inference. Crucially, RAG cannot give the model a new SKILL, only new TEXT to read. Most serious production stacks combine both. Fine-tune for behavior and house style; RAG for facts and freshness.