Jev vs Clef：Cloudflare 决策模型正面比较

import requests

JEV_ENDPOINT = "https://api.typesafe.ai/v1/jev/evaluate"
AUTO_ROUTE_CONFIDENCE = 0.85

QUESTIONS = {
    "routing": {
        "type": "choice",
        "instructions": "Which queue should this ticket go to?",
        "criteria": {
            "billing": "Payment, invoice or refund issue",
            "technical": "Product malfunction or bug",
            "sales": "Buying or upgrade question",
            "abuse": "Safety or abuse report",
        },
    },
    "needs_safety_escalation": {
        "type": "noul",
        "instructions": "Does this ticket require a safety escalation regardless of queue?",
    },
    "urgency": {
        "type": "score",
        "instructions": "Rate how urgent a human reply is, 1 (routine) to 5 (business-stopping)",
    },
}

resp = requests.post(
    JEV_ENDPOINT,
    json={"state": {"ticket": "..."}, "questions": QUESTIONS},
    timeout=5,
)
resp.raise_for_status()
data = resp.json()

route = data["routing"]  # 类型化答案 + 校准置信度，零生成
if (
    route["confidence"] >= AUTO_ROUTE_CONFIDENCE
    and not data["needs_safety_escalation"]["answer"]
):
    lane = f"auto:{route['answer']}"  # 单次前向，仅输入计费
else:
    lane = "review"  # 0.60~0.85 带宽或安全标记

import json

import requests

# 来自本站 Clef-Flash 装机实测的已验证本地路径：
# bartowski Q4_K_M GGUF、Ollama 0.35、8GB 笔记本、2048 上下文。
# 机制注记：这条路径是「生成 JSON」（21–30 个 token）——官方
# joint schema head 未被加载（427 个 backbone 张量、零
# schema head 匹配），因此没有原生概率。
OLLAMA_CHAT = "http://localhost:11434/api/chat"

SCHEMA = {
    "type": "object",
    "properties": {
        "urgent": {"type": "boolean"},
        "team": {"enum": ["billing", "technical", "sales"]},
        "severity": {"enum": ["minor", "major", "critical"]},
    },
    "required": ["urgent", "team", "severity"],
}

resp = requests.post(
    OLLAMA_CHAT,
    json={
        "model": "hf.co/bartowski/Cloudflare_clef-flash-GGUF:Q4_K_M",
        "messages": [{
            "role": "user",
            "content": "Ticket: Checkout has been failing for every "
            "customer for the last hour. Reply with urgent, team, "
            "severity as JSON.",
        }],
        "format": SCHEMA,  # 运行时层约束修的是「格式」，
        "options": {"temperature": 0, "num_ctx": 2048},  # 不是「正确性」
    },
    timeout=60,
)
reply = json.loads(resp.json()["message"]["content"])

# 热请求约 1.4~2 秒。我们的测试里，同一句宕机提示词零温度复跑
# 把 team 从 technical 翻成了 billing——路由评测要有标注样本集，
# 不能靠一张成功截图。

const JEV_ENDPOINT = "https://api.typesafe.ai/v1/jev/evaluate";
const AUTO_ROUTE_CONFIDENCE = 0.85;

const QUESTIONS = {
  routing: {
    type: "choice",
    instructions: "Which queue should this ticket go to?",
    criteria: {
      billing: "Payment, invoice or refund issue",
      technical: "Product malfunction or bug",
      sales: "Buying or upgrade question",
      abuse: "Safety or abuse report",
    },
  },
  needs_safety_escalation: {
    type: "noul",
    instructions:
      "Does this ticket require a safety escalation regardless of queue?",
  },
  urgency: {
    type: "score",
    instructions:
      "Rate how urgent a human reply is, 1 (routine) to 5 (business-stopping)",
  },
} as const;

const resp = await fetch(JEV_ENDPOINT, {
  method: "POST",
  headers: { "Content-Type": "application/json" },
  body: JSON.stringify({ state: { ticket: "..." }, questions: QUESTIONS }),
});
const data = await resp.json();

const lane =
  data.routing.confidence >= AUTO_ROUTE_CONFIDENCE &&
  !data.needs_safety_escalation.answer
    ? `auto:${data.routing.answer}`
    : "review";

// Clef 官方自称 "fully Jev-API compatible"——如果对你的调用成立，
// 换厂商就是换这个常量。保留意见：这是 Cloudflare 的自述，本站
// 未逐端点验证；而且它描述的是托管模型（本地量化路径生成 JSON、
// 无原生决策头）。把两个后端都藏在同一个接口后面，先在约 100 条
// 标注样本上验证切换，再让自动化依赖置信度数字。
const JEV_ENDPOINT = "https://api.typesafe.ai/v1/jev/evaluate";

const DECISION_ENDPOINT =
  process.env.DECISION_BACKEND === "clef"
    ? process.env.CLEF_ENDPOINT! // 你的 Workers AI / 自托管 Clef 端点
    : JEV_ENDPOINT;

const QUESTIONS = {
  routing: {
    type: "choice",
    instructions: "Which queue should this ticket go to?",
    criteria: {
      billing: "Payment, invoice or refund issue",
      technical: "Product malfunction or bug",
      sales: "Buying or upgrade question",
      abuse: "Safety or abuse report",
    },
  },
} as const;

export async function routeTicket(ticket: string) {
  const resp = await fetch(DECISION_ENDPOINT, {
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify({ state: { ticket }, questions: QUESTIONS }),
  });
  if (!resp.ok) throw new Error("decision call failed");
  const data = await resp.json();

  // 各后端的置信度语义可能不同：自动化闸门建立在「按厂商逐个
  // 验证过的校准」上，而不是「存在一个数字」上。
  return data.routing as { answer: string; confidence: number };
}