用 Clef 构建低延迟决策层：从 Workers AI 推理到强化学习微调

用户请求
   │
   ▼
Clef 决策层 ── route=direct ──> 直接处理
   │
   ├── route=tool ────────────> 调用受控工具
   │
   ├── route=large_model ─────> 转交更强模型
   │
   └── route=human ───────────> 人工审核

name = "clef-decision-router"
main = "src/index.ts"
compatibility_date = "2025-01-01"

[ai]
binding = "AI"

[vars]
# 部署前替换成 Workers AI 中实际可用的 Clef 或 Clef-flash 模型 ID
DECISION_MODEL = "REPLACE_WITH_CLEF_MODEL_ID"

interface Env {
  AI: any;
  DECISION_MODEL: string;
}

type Decision = {
  route: "direct" | "tool" | "large_model" | "human";
  confidence: number;
  reason: string;
};

const ALLOWED_ROUTES = new Set([
  "direct",
  "tool",
  "large_model",
  "human",
]);

function parseDecision(result: any): Decision {
  const raw =
    typeof result === "string"
      ? result
      : result?.response ?? result?.output ?? "";

  const parsed = JSON.parse(raw);

  if (!ALLOWED_ROUTES.has(parsed.route)) {
    throw new Error("Model returned an unsupported route");
  }

  const confidence = Number(parsed.confidence);
  if (!Number.isFinite(confidence) || confidence < 0 || confidence > 1) {
    throw new Error("Invalid confidence score");
  }

  return {
    route: parsed.route,
    confidence,
    reason: String(parsed.reason ?? ""),
  };
}

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    if (request.method !== "POST") {
      return new Response("Use POST", { status: 405 });
    }

    const body = (await request.json()) as { message?: string };
    const message = body.message?.trim();

    if (!message) {
      return Response.json({ error: "message is required" }, { status: 400 });
    }

    const prompt = `You are a request router.
Choose exactly one route:
- direct: simple question that can be answered without tools
- tool: requires an approved internal tool
- large_model: requires deeper reasoning
- human: security, payment, privacy, or irreversible action

Return JSON only:
{"route":"direct|tool|large_model|human","confidence":0.0,"reason":"short reason"}

Request:
${message}`;

    try {
      const result = await env.AI.run(env.DECISION_MODEL, { prompt });
      let decision = parseDecision(result);

      // 低置信度时采取保守策略，而不是盲目信任模型。
      if (decision.confidence < 0.65) {
        decision = {
          route: "human",
          confidence: decision.confidence,
          reason: `Low-confidence fallback: ${decision.reason}`,
        };
      }

      return Response.json(decision);
    } catch (error) {
      return Response.json(
        {
          route: "human",
          confidence: 0,
          reason: "Decision model failed; using safe fallback",
        },
        { status: 200 }
      );
    }
  },
};

npm install --save-dev wrangler
npx wrangler dev

curl -s http://localhost:8787 \
  -H 'content-type: application/json' \
  -d '{"message":"请删除我的账户以及全部历史订单"}'

{"input":"查询昨天的快递状态","action":"tool","reward":1.0,"metadata":{"tool":"shipment_lookup"}}
{"input":"帮我转账 5000 元到新账户","action":"human","reward":1.0,"metadata":{"risk":"financial"}}
{"input":"解释一下什么是幂等性","action":"direct","reward":1.0,"metadata":{"topic":"engineering"}}
{"input":"分析这份包含多个附件的合同","action":"large_model","reward":0.8,"metadata":{"complexity":"high"}}

import json
import sys

ALLOWED_ACTIONS = {"direct", "tool", "large_model", "human"}


def validate(path: str) -> None:
    total = 0
    action_counts = {action: 0 for action in ALLOWED_ACTIONS}

    with open(path, "r", encoding="utf-8") as file:
        for line_number, line in enumerate(file, start=1):
            if not line.strip():
                continue

            row = json.loads(line)
            total += 1

            if not isinstance(row.get("input"), str) or not row["input"].strip():
                raise ValueError(f"line {line_number}: invalid input")

            action = row.get("action")
            if action not in ALLOWED_ACTIONS:
                raise ValueError(f"line {line_number}: invalid action {action!r}")

            reward = row.get("reward")
            if not isinstance(reward, (int, float)) or not -1 <= reward <= 1:
                raise ValueError(f"line {line_number}: reward must be between -1 and 1")

            action_counts[action] += 1

    if total == 0:
        raise ValueError("dataset is empty")

    print(f"valid rows: {total}")
    for action, count in sorted(action_counts.items()):
        print(f"{action}: {count} ({count / total:.1%})")


if __name__ == "__main__":
    if len(sys.argv) != 2:
        raise SystemExit("usage: python validate_dataset.py decisions.jsonl")
    validate(sys.argv[1])

python validate_dataset.py decisions.jsonl