Fine-tuning sounds serious. Retrieval-augmented generation (RAG) sounds like a research paper. Most commercial software needs the second, a small amount of the first, and a lot of boring engineering in between.
What RAG is for
Your policies, catalog, tickets, and runbooks change weekly. You do not want to retrain a model every time legal edits a paragraph. RAG keeps the source of truth in a store you control, embeds it, and stuffs the relevant chunks into the prompt. The model’s job is to speak clearly, not to remember your Q3 pricing. That is the pattern we use most on AI engagements.
When fine-tuning earns its keep
Tone that must match a brand across thousands of generations. Classification that a prompt cannot stabilize. A format so strict that few-shot examples keep falling over. Even then, fine-tune on tasks — not on “all our PDFs.” Documents belong in retrieval. Weights belong to behaviour.
The part vendors skip
Chunking, access control, freshness, and evals. A RAG system that retrieves the wrong customer’s contract is worse than no AI. We treat retrieval like any other query: indexes, tests, and an audit trail. If you are already on AWS, the plumbing sits next to the rest of the environment — not in a separate “AI cloud” nobody owns.
Need a partner who will argue for the thinner option? That is usually the right first conversation.