Concepts 6 minReviewed September 1, 2026
RAG vs fine-tuning vs LoRA
Decide whether the model needs new facts, new behaviour, or both.
The answer
Use RAG to supply changing facts at request time. Use LoRA or another parameter-efficient method to change repeatable behaviour, style, or task performance. Combine them when the model needs both.
Use RAG for knowledge
- Policies and catalogues that change.
- Answers that must cite a source.
- Private documents that should not be baked into weights.
Use LoRA for behaviour
- A stable response format or tone.
- A narrow classification or extraction task.
- A repeatable capability that prompting does not teach reliably.
Do not fine-tune first
Begin with a strong evaluation set and prompt baseline. If failures are missing facts, improve retrieval. If the model sees the right evidence but behaves incorrectly, then test a parameter-efficient adaptation.