RAG vs fine-tuning
Two ways to give a model your knowledge.
Two ways to hand a model your knowledge: one keeps it on a shelf you can reach, the other writes it into muscle memory.
When people say they want a model to “know their business”, they usually mean one of two very different things. Confusing the two is the most common, and most expensive, mistake in applied AI.
Retrieval keeps facts fresh
Retrieval, or RAG, leaves your documents in a searchable store and, at the moment a question arrives, hands the model only the passages that matter. Facts stay current and editable: change the document, change the answer. Nothing is baked in.
Retrieval for what is true today. Fine-tuning for how the answer should sound.
Fine-tuning teaches a manner
Fine-tuning instead writes patterns directly into the model’s weights. It is the right tool for style, format, and skill, for teaching a model how to behave, and the wrong tool for raw facts, which it will memorise and then slowly get wrong.
Most systems worth running use both: retrieval for the truth, fine-tuning for the voice. The art is in knowing which question you are actually asking.