An online retailer's Foundry-hosted customer service assistant must always reflect the current return policy, which the operations team revises nearly every week. The AI team wants the assistant to pick up each revision automatically as soon as the policy document is updated, without retraining or redeploying the model. Which approach should they use?
Select an answer to reveal the explanation.
Short Explanation
Think of the policy document as something that changes almost every week, and the goal is for the assistant to always be quoting from this week's version without anyone having to rebuild the model. The way to do that is to hook the assistant up to a search index over the live document, so every time someone asks a question, it goes and looks up whatever is currently in that document and answers from there. Retraining the model on the new wording every time it changes would work, technically, but it means a whole training cycle every single week just to keep up with a document edit, which is way more effort than refreshing an index. Turning up the randomness dial does not make the assistant more current, it just makes the wording less predictable. And trying to paste every past version of the policy into the prompt as examples just clutters things up and does not solve the freshness problem, since nobody is guaranteed to remember to add the newest version in time.
Full Explanation
The correct answer is B. Connecting the assistant to a retrieval-augmented generation index built over the live policy document means every answer is generated from whatever version of the policy is currently indexed, so a weekly revision to the source document flows through to the assistant the next time the index refreshes, with no model retraining or redeployment required. Option A is incorrect because fine-tuning bakes a snapshot of the policy into the model weights; doing that every week means a new training run, evaluation, and deployment cycle each time, which is slow and expensive compared to simply updating an index. Option C is incorrect because temperature controls how varied or creative the wording of a response is, not which facts the model has access to; raising it does nothing to keep the assistant current and can make responses less predictable. Option D is incorrect because stuffing the full history of policy revisions into the system message as few-shot examples bloats the context with outdated versions, does not automatically pick up new updates without manually editing the prompt, and risks confusing the model about which version is current. Grounding through retrieval is the mechanism designed specifically for fast-changing source material.