A model must answer from frequently updated private operating procedures without retraining whenever a document changes. Which architecture is most appropriate?

1. Increase temperature to invent missing procedures
2. Fine-tune once and assume future policy changes are included
3. Retrieval-augmented generation with an authorized current knowledge index
4. Place every document in every prompt regardless of limits