Should you use RAG or fine-tuning to teach a model new information?

RAG for facts that change or need a citation, fine-tuning for a fixed skill, tone, or format you want baked into the model itself. RAG leaves the model’s weights untouched and looks up the relevant chunk of your corpus at request time, so updating a document takes effect on the next query with no retraining, and it gives you something fine-tuning can’t: the actual passage an answer drew from, so a reader can check it. Fine-tuning does the opposite. It changes what the model does by adjusting its weights, so the new behavior holds even without a retrieval step, but every change to what it should know means training again.

The two aren’t a choice between opposites so much as a division of labor. A fine-tuned model can still receive retrieved context, and teams commonly fine-tune the stable parts of behavior, like tone, output format, or when to call a tool, while using RAG for whatever keeps changing underneath it, like documentation, prices, or policy.