Fine-tuning vs RAG

Result

Suggested: Retrieval (RAG). Scores: Retrieval (RAG) 15, Prompting alone (no RAG, no fine-tuning) 3, Fine-tuning -1. Guidance from a stated rubric, not a test result.

Retrieval (RAG)

15

Prompting alone (no RAG, no fine-tuning)

3

Fine-tuning

-1

How each answer scored

• Facts from our own documents or dataRetrieval +3Retrieval puts the source text in the prompt; fine-tuning is an unreliable way to store facts.
• Every few monthsRetrieval +2, Fine-tuning +0.5Retrieval still updates more easily than a training run.
• YesRetrieval +3Retrieval knows which document each passage came from, so it can cite it. A fine-tuned model cannot.
• A few dozenPrompting alone +2, Fine-tuning +0.5A few dozen fit as examples in the prompt; often too few to fine-tune well.
• Far more than one promptRetrieval +3Retrieval picks the few relevant passages from a large corpus.
• Yes, permissions differRetrieval +3Retrieval can filter by permission per query. Anything trained into a model is visible to every user of it.
• Normalno pointsNo effect.
• Not yetRetrieval +1, Prompting alone +1, Fine-tuning -2Without that capacity a fine-tuned model goes stale; retrieval and prompting are easier to maintain.

Guidance from a transparent scoring rubric, not a guarantee. The points per answer are planning assumptions; test the approach on your own data.