Jev vs Clef: Cloudflare’s Decision Models Compared

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"]  # typed answer + calibrated confidence, no generation
if (
    route["confidence"] >= AUTO_ROUTE_CONFIDENCE
    and not data["needs_safety_escalation"]["answer"]
):
    lane = f"auto:{route['answer']}"  # single pass, input tokens only
else:
    lane = "review"  # 0.60-0.85 band or safety flag

import json

import requests

# The verified local path from our Clef-Flash install test:
# bartowski Q4_K_M GGUF in Ollama 0.35, 8 GB laptop, 2,048 ctx.
# Mechanism note: this path GENERATES JSON (21-30 tokens) - the
# official joint schema head is not loaded (427 backbone tensors,
# zero schema-head matches), so there are no native probabilities.
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,  # runtime-level constraint fixes the FORMAT,
        "options": {"temperature": 0, "num_ctx": 2048},  # not correctness
    },
    timeout=60,
)
reply = json.loads(resp.json()["message"]["content"])

# Warm requests: ~1.4-2s. A zero-temperature repeat of the same outage
# prompt flipped team from "technical" to "billing" in our test -
# evaluate routing on a labeled set, not on one screenshot.

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 officially describes itself as "fully Jev-API compatible" -
// if that holds for your calls, switching vendors is this constant.
// Caveats: it is Cloudflare’s self-description, unverified by this
// site endpoint by endpoint, and it describes the hosted model (the
// local quantized path generates JSON without a native head). Keep
// both backends behind one interface and validate the swap on ~100
// labeled examples before automating on the confidence number.
const JEV_ENDPOINT = "https://api.typesafe.ai/v1/jev/evaluate";

const DECISION_ENDPOINT =
  process.env.DECISION_BACKEND === "clef"
    ? process.env.CLEF_ENDPOINT! // your Workers AI / self-hosted Clef endpoint
    : 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();

  // Confidence semantics may differ per backend: gate automation on
  // calibration validated per vendor, not on the existence of a number.
  return data.routing as { answer: string; confidence: number };
}