Jev 判断模型

Jev 由 TypeSafe 提供，是判断模型，不是聊天模型。给它一段文字和一个问题，它直接回答归哪一类、打几分、是还是否，结果可以被程序直接使用。典型场景：把"订单重复扣款"的客户反馈自动分到账单团队、给一条评论打满意度分、判断一封邮件是否在要求退款。

请求示例

下面把一条"重复扣款"的反馈分到 billing、technical 或 other。模型名固定为 Jev。

curl "https://apicdn.xyc.ai/v1/systemone" \
  -H "Authorization: Bearer sk-xxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Jev",
    "state": "I was charged twice for the same order.",
    "questions": {
      "department": {
        "type": "choice",
        "instructions": "Which team should handle this customer message?",
        "criteria": {
          "billing": "Charges, payments and refunds",
          "technical": "Bugs and integration issues",
          "other": "Other requests"
        }
      }
    }
  }'

响应结构

这条反馈被分到 billing，并返回各分类的概率。answers.department.choice 是选中的分类，probabilities 是各分类概率，confidence 是模型对这次判断的置信度。

{
  "model": "Jev",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "probabilities": {"billing": 0.96, "technical": 0.01, "other": 0.03},
      "confidence": 0.92
    }
  },
  "usage": {"input_tokens": 339, "output_tokens": 38}
}

三种判断类型

state 可以是文本、对象或数组。questions 里每个自定义问题名都会在 answers 中对应返回；多个问题分别独立判断，前一个问题的答案不会传给后一个。

评分示例

{
  "model": "Jev",
  "state": "这个产品用了三天就坏了，客服态度也很差。",
  "questions": {
    "sentiment": {
      "type": "score",
      "instructions": "评价这条评论的满意度",
      "criteria": ["非常不满", "不满", "一般", "满意", "非常满意"]
    }
  }
}
// answers.sentiment → {"type":"score","score":0.01,"legend":{"0":"非常不满",...},"probabilities":{"0":0.99,...},"confidence":0.99}

是非示例

{
  "model": "Jev",
  "state": "Hi, I want to cancel my subscription and get a refund.",
  "questions": {
    "is_refund": {"type": "noul", "instructions": "Is the customer asking for a refund?"}
  }
}
// answers.is_refund → {"type":"noul","noul":0.99}

用官方 SDK 接入

TypeSafe 官方 SDK 可直接使用，只需把 base_url 指向网关、api_key 换成 XycAi 密钥，模型指定为 Jev。

Python（pip install typesafe-sdk）：

import os
from typesafe_sdk import Choice, TypeSafeClient

client = TypeSafeClient(
    api_key=os.environ["XYCAI_API_KEY"],
    base_url="https://apicdn.xyc.ai",
    model="Jev",
)
response = client.system_one(
    state="I was charged twice for the same order.",
    questions={
        "department": Choice(
            instructions="Which team should handle this customer message?",
            criteria={
                "billing": "Charges, payments and refunds",
                "technical": "Bugs and integration issues",
                "other": "Other requests",
            },
        ),
    },
)
print(response.answers["department"])

TypeScript（npm install @typesafe-ai/sdk）：

import { choice, TypeSafeClient } from "@typesafe-ai/sdk";

const client = new TypeSafeClient({
  apiKey: process.env.XYCAI_API_KEY,
  baseURL: "https://apicdn.xyc.ai",
});
const response = await client.systemOne({
  model: "Jev",
  state: "I was charged twice for the same order.",
  questions: {
    department: choice("Which team should handle this customer message?", {
      billing: "Charges, payments and refunds",
      technical: "Bugs and integration issues",
      other: "Other requests",
    }),
  },
});
console.log(response.answers.department);

价格与限制

按 输入 token 计费，输出免费；以模型广场 TypeSafe 企业版实时价格为准。state 加最长的单个问题不超过 32K tokens，state 加全部问题合计不超过 64K tokens。

排查调用问题

接口定义参考 TypeSafe API 文档，三种类型的完整写法见 TypeSafe primitives。