A public-health agency wants a GenAI assistant to answer resident questions about current health advisories accurately, and the advisories change on a weekly basis. At a conceptual level, weighing fine-tuning against retrieval-augmented generation, which approach should the agency choose to keep answers current, and why?
Select an answer to reveal the explanation.
Short Explanation
Think about a reference librarian versus a memorized speech: fine-tuning is the speech, locked in until someone rehearses it all over again, while retrieval-augmented generation is the librarian who walks over to the current shelf before answering. When advisories change weekly, you want the librarian, not the speech.
Full Explanation
Retrieval-augmented generation connects a GenAI assistant to a live content source, so each answer is grounded in whatever document is current at the moment the question is asked, without touching the model itself. That mechanism is exactly what a weekly-changing advisory calendar needs: the underlying model stays fixed while the retrieved content refreshes on its own schedule. Fine-tuning fails this scenario by concept, not by execution quality: it bakes a snapshot of facts into the model's weights, so as soon as the next advisory is published, the model quietly starts answering from stale information until someone runs a new training cycle. Repeating that retraining every time a policy changes turns a documentation update into an engineering project, which is the wrong tool for the pace of change described. Treating retrieval as a temporary launch step and fine-tuning as the long-term destination gets the tradeoff backwards, since fine-tuning does not get cheaper to maintain as update frequency rises — it gets more expensive. A caveat worth remembering: fine-tuning still has a real place for teaching a model tone, format, or domain-specific style, just not for facts that expire quickly. As an operational check, ask what the update cadence is for the underlying content; if it is measured in days or weeks, retrieval belongs in the design.