A financial services firm wants its Foundry-hosted assistant to answer employee questions about the exact figures in this quarter's internal earnings report, which is replaced with new numbers every three months. The team is deciding how to give the model access to these figures so answers stay accurate as the report changes each quarter, without retraining a model every time. Which approach best fits this need?
Select an answer to reveal the explanation.
Short Explanation
The tricky part of this situation is that the source of truth changes every few months, so whatever the assistant relies on can't be locked in once and left alone. Baking this quarter's numbers into the model itself through extra training would work for a little while, but the moment a new report comes out, those numbers are outdated and the whole training process has to happen again just to catch up. That's a lot of ongoing effort for something that should just reflect whatever document is currently in force. The better move is to let the assistant look up the current report each time someone asks a question, rather than trying to remember it. That way, updating the assistant for a new quarter is as simple as swapping in the new document, no retraining required. The other two options, letting it write longer answers or making its wording more unpredictable, don't touch the actual issue, which is where the numbers come from in the first place.
Full Explanation
The correct answer is C. Retrieval-augmented generation lets the assistant pull the actual text of the current quarter's report from an index at the moment a question is asked, so the model answers from the live document rather than from what it happened to memorize during training. When the report changes next quarter, the team only needs to update the index, not retrain anything. Option A is incorrect because fine-tuning on this quarter's numbers bakes them into model weights that go stale the moment a new report is issued, requiring a costly retraining cycle every three months just to stay current. Option B is incorrect because output token limits control how long a response can be, not whether the model has access to accurate source figures; a longer response built on a model's guesswork is still inaccurate. Option D is incorrect because raising temperature increases the randomness and creativity of wording, which works against precise numeric recall rather than improving it, and could make the assistant less consistent on exactly the figures that must be exact. Grounding the assistant in the current report through retrieval is the approach that keeps answers accurate as the source data changes on its own schedule.