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。

相關引導路徑保持聊天機器人專注
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。

  1. 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
  2. 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
  3. Which control acts on retrieved document chunks before they enter model context?
    • Tool authorization
    • Post-release analytics
    • Retrieval check
    • Output check
  4. 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
  5. 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
  6. 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
  7. 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
  8. 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