引導式測驗 · 硬 等級
保持聊天機器人專注的小測驗
測試範圍陳述、檢索邊界、分階段控制、工具授權、後備行為以及主題檢查無法證明正確性的原因。
問題 1 的 8
What should a useful chatbot scope statement make clear?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What should a useful chatbot scope statement make clear?
- 僅型號名稱
- Topics, permitted sources and tasks, and when to refuse or hand off
- A promise that every answer is accurate
- 模型中參數的數量
- How can a team reduce unsupported answers when its approved corpus has no relevant material?
- Hide retrieval results from evaluation
- Require the assistant to guess
- Provide a “not covered” response or a handoff
- Expand tool permissions automatically
- Which control acts on retrieved document chunks before they enter model context?
- Tool authorization
- Post-release analytics
- Retrieval check
- Output check
- Why is a prompt that says “stay on topic” insufficient by itself?
- It does not enforce data, retrieval, or tool boundaries and cannot guarantee every answer follows it
- It prevents all user questions
- It disables the model
- Prompts cannot communicate any policy
- What should happen before a chatbot executes an external action?
- Trust that the answer stayed on topic
- Validate the tool and its arguments against application permissions
- Check only the final display text
- Retrieve more documents
- What does a question-to-corpus similarity threshold establish?
- That no human handoff is needed
- That the answer is correct
- That a request is legally safe
- A routing signal, not proof that the response is grounded or correct
- Which evaluation set best tests whether scope controls work?
- Only spelling errors
- Only model benchmark prompts unrelated to the service
- Only easy in-scope requests
- In-scope, borderline, unsupported, adversarial, and outage scenarios
- Why should teams track both inappropriate answers and unnecessary refusals?
- Only refusal rate matters
- A perfect classifier never has tradeoffs
- Refusals are unrelated to scope
- A control can miss out-of-scope requests or block valid edge cases