A telecom wants a chatbot built on a hosted Gemini model to discuss only its plans and devices, and to decline everything else courteously. Which configuration is the most practical?
Show answer and explanation
Answer: D
A system prompt in Vertex AI sets the assistant's role, permitted subjects and how to respond when a request falls outside them. Pairing that with grounding in the plan catalog keeps factual answers tied to approved content. This achieves the scope control without training cost, and it is quick to revise as plans change.
Why the other options are wrong
- A. Fine-tuning can shape tone but is slower and costlier to maintain, and it does not reliably enforce refusals on new off-topic questions the way explicit instructions do.
- B. A separate classifier adds a second model to build, monitor and keep current, when the system instructions can already handle scope. It is not the most practical starting point.
- C. A token cap truncates text mid-sentence without deciding what is on topic. The bot could still produce off-topic content within the limit, so length does not enforce scope.