RAG or fine-tuning: how to tell which one your AI model needs

When an AI model in a product errs, the cause is usually either missing knowledge or correct knowledge applied with the wrong format, tone, or decisions. Retrieval-augmented generation (RAG) addresses the first problem by supplying relevant documents at answer time, leaving the model itself unchanged. Fine-tuning instead retrains the model's internal weights on examples of the task, which suits narrow, repetitive jobs with abundant labeled data. RAG is preferred when information changes often, is private or too large to fit in a prompt, or must be cited, while fine-tuning fits behavior problems, latency or cost reduction, and tasks with real examples. The article notes that each approach has limits, and that fine-tuning is a poor method for adding facts.