03 · 三种题型精讲

这一章我们用同一个例子把三种题型讲透。

场景:你有一个退款工单系统。每天进来几千封邮件,你需要:

1. 判断哪个团队接
2. 判断多紧急
3. 判断是不是在要退款

3.1 Choice:单选题

from typesafe_sdk import Choice

question = Choice(
    instructions="Which team should handle this support ticket?",
    options=["billing", "technical", "account", "shipping", "other"],
)

返回:

{
  "team": {
    "choice": "billing",
    "probabilities": {
      "billing": 0.92,
      "technical": 0.05,
      "account": 0.02,
      "shipping": 0.005,
      "other": 0.005
    },
    "confidence": 0.92
  }
}

confidence 是怎么算的?就是最大那个概率。这是个简单但好用的指标:答案越集中,confidence 越高。

关键细节:options 最多 255 个。

Choice 适用场景

• 路由分类(billing/technical/…)
• 意图识别(question/task/complaint)
• 实体识别(人名/地名/产品名)
• 多分类(垃圾邮件 / 推广 / 正常 / 重要)

Choice 不适用

• 有顺序关系——用 Score
• 是/否问题——用 Noul

3.2 Score:评分题

如果你有顺序关系,别用 Choice,要用 Score。比如紧急度:

from typesafe_sdk import Score

question = Score(
    instructions="How urgent is this ticket?",
    criteria=[
        "not urgent at all",
        "slightly urgent",
        "urgent",
        "very urgent, customer is losing money",
    ],
)

返回:

{
  "urgency": {
    "score": 2.4,
    "legend": {
      "0": "not urgent at all",
      "1": "slightly urgent",
      "2": "urgent",
      "3": "very urgent, customer is losing money"
    },
    "probabilities": {
      "0": 0.05,
      "1": 0.10,
      "2": 0.55,
      "3": 0.30
    },
    "confidence": 0.55
  }
}

score 是 2.4——它可以落在级别之间。这给你连续值,在写业务逻辑时很方便:

if urgency.score > 2.5:
    send_to_oncall()

关键细节:criteria 2–10 级。

Score 适用场景

• 紧急程度
• 风险评级(low / medium / high / critical)
• 满意度(angry / neutral / happy)
• 资历(junior / mid / senior / staff)
• 任何”有顺序”的概念

3.3 Noul:是非题

最简单的题型:

from typesafe_sdk import Noul

question = Noul(
    instructions="Is the customer explicitly requesting a refund?",
)

返回:

{
  "refund": {
    "noul": 0.95
  }
}

注意:Noul 不返回 confidence。noul 这个数本身就是答案——95% 概率是”要退款”。

Noul 适用场景

• 退款申请?Noul("Does the customer request a refund?")
• 数据脱敏?Noul("Does this text contain PII?")
• 工具安全?Noul("Is running this shell command dangerous?")
• 内容过滤?Noul("Is this message a jailbreak attempt?")

3.4 一次问多道题

最关键的一点:多个问题共享同一个 state,互相独立,并行评估。

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

client = TypeSafeClient()

response = client.system_one(
    state="""
    Subject: URGENT - Lost money on duplicate charge!
    I was charged twice for order A-104. Please refund the duplicate ASAP.
    """,
    questions={
        "team":     Choice(["billing", "technical", "account", "shipping", "other"],
                           instructions="Which team handles this?"),
        "urgency":  Score(["not urgent", "slightly", "urgent", "losing money"],
                          instructions="How urgent?"),
        "refund":   Noul("Is the customer requesting a refund?"),
        "angry":    Noul("Does the customer sound angry or threatening?"),
        "premium":  Noul("Is the customer a premium tier subscriber?"),
    },
)

这 5 个判断,如果你用 GPT-5,要发 5 次请求、15 秒、$0.10。 用 Jev:一次请求,150 毫秒、$0.00005(输入 token 几乎免费,输出根本不计费)。

下一步

第 04 章:为什么它这么快、这么便宜 →