When is RAG a better fit than fine-tuning?

Example Answer

I start with RAG when the product promise depends on changing facts, source-specific answers, permissions, or citations. If users need the latest policy, the right contract clause, or a traceable answer tied to a document, retrieval is the better primitive.

Fine-tuning is better when the gap is behavior: response style, output structure, tool habits, or domain phrasing. It is a weak fit when the real problem is freshness or access to private knowledge. I do not want to retrain a model every time the corpus changes.

The tradeoff is operational complexity. RAG adds ingestion, indexing, access control, and observability work. I accept that complexity when it buys me auditability and the ability to fail closed. In practice, if I need current evidence and I want to debug the system layer by layer, I pick retrieval first and only consider fine-tuning later if behavior is still the bottleneck.

Common Poor Answer

A weak answer is saying, "RAG is for knowledge and fine-tuning is for style," and stopping there. That misses freshness, citations, permissions, and the operational tradeoff that actually drives the architectural choice.