Use retrieval when knowledge changes.
Retrieval-augmented generation (RAG) searches approved sources and supplies relevant material to a language model for the current request. That makes it useful for policies, product information, technical documentation, customer history or other knowledge that should remain separable from the base model.
Use fine-tuning for behavior patterns when justified.
Fine-tuning can help with specialized output patterns, task behavior or domain examples, but it is not a convenient database for frequently changing facts. It also introduces its own data preparation, evaluation and model lifecycle requirements.
| Requirement | RAG | Fine-tuning |
|---|---|---|
| Frequently changing facts | Often appropriate | Usually not the first choice |
| Source citations | Can support them | Does not inherently provide provenance |
| Change output style/behavior | Limited | Potentially appropriate |
| Permission-aware knowledge | Can be designed into retrieval | Harder to express as dynamic source permissions |
Evaluation matters more than the label.
A RAG system can still retrieve the wrong passage, and a fine-tuned model can still fail the actual workflow. Test the architecture against representative questions, expected actions and failure cases before choosing complexity for its own sake.
Related service: Data & Knowledge Intelligence.
References and further reading
The original retrieval-augmented generation research is described in Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”. For a current example of supervised fine-tuning as task adaptation, see the OpenAI fine-tuning guide.