置信度门控路由

置信度门控路由

把置信度当作第二条轴。答案告诉你该做什么；置信度告诉你该不该做。

TypeSafe 最强大的特性之一，是置信度。只要你有意识地把决策门控在置信度上，就能构建出既可靠又安全的系统。

示例：语音银行指令

设想你要做一个语音银行界面，让用户能口头操作自己的账户。解读用户意图时，你当然希望始终有说得过去的置信度，但有些操作比另一些风险更高，因而需要更高的置信度阈值。

%%{init: {"fontFamily": "Inter, sans-serif", "flowchart": {"rankSpacing": 35, "wrappingWidth": 300, "subGraphTitleMargin": {"top": 12, "bottom": 36}}}}%%
flowchart LR
    command["voice banking command"]

    subgraph req["TypeSafe evaluates<br/>the question"]
        intent["<b>Choice:</b> intent"]
    end

    command -- "one request<br/>command + intent<br/>question" --> req
    req -- "one response<br/>intent answer +<br/>confidence" --> gate{"<b>confidence high enough?</b><br/>your code"}
    gate -- "below 0.6<br/>or other intent" --> human["send to a support agent"]
    gate -- "check_balance<br/>at least 0.6" --> balance["show the balance"]
    gate -- "approve_transfer<br/>0.6 to 0.85" --> confirm["ask the user to confirm"]
    gate -- "approve_transfer<br/>above 0.85" --> approve["approve the transfer"]

第 1 步：判断用户意图

{
  "intent": {
    "type": "choice",
    "instructions": "What action is the user requesting?",
    "criteria": {
      "check_balance": "Check the balance of an account",
      "approve_transfer": "Approve the pending transfer request",
      "other": "Something else"
    }
  }
}

第 2 步：置信度门控路由

action = response.answers["intent"]

# Below 0.6 confidence on any action, route to a human
if action.confidence < 0.6:
    route_to_support_agent(account_id)

elif action.choice == "check_balance":
    # Low stakes. 0.6 confidence is sufficient.
    show_balance(account_id)

elif action.choice == "approve_transfer":
    if action.confidence > 0.85:
        # High stakes, but high confidence. Safe to act automatically.
        approve_transfer(account_id)
    else:
        # High stakes, moderate confidence. Verify intent first.
        ask_user_to_confirm("Just to confirm: you would like to approve this transfer, is that correct?")

else:
    route_to_support_agent(account_id)

0.6 这条底线，兜住的是模型确实拿不准的部分。底线之上，每种操作各自有阈值，取决于分类错了还要照做会有什么后果。以 0.6 的置信度去查余额没问题，因为最坏的情况不过是用户听一遍余额播报。但批准一笔转账需要很高的置信度（>0.85），否则系统就该请用户先确认一下。

想更详细地了解在系统里该如何看待置信度，见置信度。