A support assistant must answer from product manuals updated daily and show the source passages. The model already follows the desired response format. Which approach best provides freshness and provenance with minimal retraining?

1. Increase temperature so the model invents more recent details.
2. Fine-tune weekly and assume model weights can return exact document provenance.
3. Use retrieval-augmented generation with a refreshed knowledge source and returned citations.
4. Keep one static prompt containing last year manuals and disable retrieval.