A furniture retailer deploys a Foundry-hosted chat assistant on its website to answer questions about product dimensions and delivery. During testing, customers ask the assistant to write poems and discuss unrelated topics, and it happily complies, drifting from its intended purpose. What should the team configure to keep the assistant focused on furniture-related questions and declining unrelated requests?
Select an answer to reveal the explanation.
Short Explanation
The assistant is not being unruly on purpose — nobody ever told it what it is for. A system message is where you spell that out up front: you are a furniture help assistant, answer questions about dimensions and delivery, and politely decline anything else. That single instruction, set before the conversation even starts, is what keeps it on task. Trimming the reply length just cuts a poem off partway through instead of preventing it. Turning up the randomness dial makes the assistant less predictable, which is the opposite of what you want here. And swapping to a smaller, cheaper model does not automatically teach it to stay in its lane — without that same role-and-boundaries instruction, a smaller model can wander off topic just as easily as a bigger one.
Full Explanation
The correct answer is D. A system message sets the assistant's role, scope, and behavioral boundaries before any user turn begins, so instructing it to stay focused on furniture dimensions and delivery, and to decline unrelated requests such as writing poems, directly addresses the drift the team observed in testing. Option A is incorrect because lowering the maximum output token limit only shortens how long a response can be; it does not stop the assistant from attempting to engage with an off-topic request, it just cuts the reply short. Option B is incorrect because increasing temperature makes output more varied and unpredictable, which would make an already off-topic assistant even harder to control, not more focused. Option C is incorrect because switching to a smaller model changes general capability and cost, but a smaller model has no inherent instruction to refuse unrelated topics — without a system message defining scope, it could drift just as easily as the larger one.