Fine-tuning vs RAG

Two answers to two different questions

Fine-tuning adjusts the model's weights on examples, teaching it a format, a style or the shape of a task. Retrieval leaves the model alone and puts the relevant facts into the request.

They solve different problems. Fine-tuning is bad at knowledge: what it learns is frozen at training time, and a changed fact means training again. Retrieval is bad at behavior: no amount of context teaches a model to reliably produce your output format.

The useful diagnostic is one question. “It does not know about our products” is a retrieval problem. “It will not answer the way we need” is a fine-tuning problem. Most teams reach for the second when they have the first, and spend a month finding out.