Why use RAG rather than fine-tuning to answer from company policies that change often?
Why A is the answer
Retrieval uses the current data at query time
Updating an index is simple, and the model sees current content. RAG reduces, but does not eliminate, wrong answers.
In detail
- Retrieval uses the current data at query time.
- Fine-tuning fixes knowledge at training time.
- RAG can still be wrong, so test it.
Use case
An illustrative scenario, not a real incident.
Company policies change monthly. RAG reads the current text at query time, so an update applies at once, with no retraining.
See it in different ways
sequenceDiagram participant U as User participant A as App participant S as Search index participant M as Model U->>A: Question A->>S: Search S-->>A: Relevant passages A->>M: Question with passages M-->>U: Grounded answer
sequenceDiagram participant Q as Query participant K as Keyword search participant V as Vector search participant R as Ranking Q->>K: Search Q->>V: Search K->>R: Ranked results V->>R: Ranked results R->>R: Reciprocal Rank Fusion R-->>Q: Merged, then semantic ranker
flowchart TD
Q{"What do you need?"}
Q -->|"Facts that change"| A0["RAG"]
Q -->|"A stable style"| A1["Fine-tuning"]Key point: RAG uses current data at query time.
Why the other options are wrong
- B. RAG removes the need for any language model
- RAG still uses a language model.
- C. RAG guarantees that every answer will be correct
- RAG cannot guarantee correct answers.
- D. Fine-tuning always costs nothing, but it is slower to run
- Fine-tuning is not free.