Jev 决策模型:使用教程与多行业应用场景
Jev 是 TypeSafe AI 的 System One(系统 1)决策模型。它不生成自由文本,只输出结构化判定结果:一次请求提交一份
state上下文和一组questions,并行返回概率、选项与分数,可直接喂给代码里的if。
一、Jev 是什么
|
对话模型 |
Jev | |
|---|---|---|
|
输出 |
自由文本 |
结构化判定(概率 / 选项 / 分数) |
|
交互 |
一问一答 |
一次提交多组问题,并行独立判定 |
|
消费方式 |
还得解析文本 |
直接取值 |
|
计费 |
输入 + 输出 |
输入 token |
对话模型负责写,Jev 负责判。它适合的不是「生成」,而是「判断」——分类、路由、打分、分流、校验。
二、一分钟上手
端点
POST https://api.new.bi/v1/systemone
Authorization: Bearer sk-xxxxxxxx
Content-Type: application/json
请求体只要三个字段:model、state、questions。本文示例统一使用 https://api.new.bi 作为接入地址。
⚠️ 请使用
/v1/systemone。换成/v1/chat/completions会导致 token 统计与上游账单严重不一致。
curl
curl -X POST 'https://api.new.bi/v1/systemone' \
-H "Authorization: Bearer sk-xxxxxxxx" \
-H 'Content-Type: application/json' \
-d '{
"model": "jev-1.13.0",
"state": "用户反馈:结账页面点击支付后一直空白,换了两个浏览器都不行,很生气",
"questions": {
"is_bug": {
"type": "noul",
"instructions": "用户是否在上报产品缺陷"
},
"team": {
"type": "choice",
"instructions": "该工单应交给哪个团队",
"criteria": {
"billing": "账单、发票、退款",
"tech": "功能缺陷、集成故障",
"account": "账号与资料问题"
}
},
"fury": {
"type": "score",
"instructions": "客户情绪严重程度",
"criteria": ["平静", "沮丧", "非常愤怒"]
}
}
}'
真实返回:
{
"model": "jev-1.13.0",
"answers": {
"fury": {
"type": "score",
"score": 1.73,
"confidence": 0.6,
"legend": { "0": "平静", "1": "沮丧", "2": "非常愤怒" },
"probabilities": { "0": 0.0, "1": 0.27, "2": 0.73 }
},
"is_bug": { "type": "noul", "noul": 0.97 },
"team": {
"type": "choice",
"choice": "tech",
"confidence": 0.27,
"probabilities": { "billing": 0.48, "tech": 0.52, "account": 0.0 }
}
},
"usage": { "input_tokens": 409, "output_tokens": 69 }
}
一次请求,三个判定,全是可直接用的结构。
Python
import requests
def jev(state, questions, *, base, key, model="jev-1.13.0"):
r = requests.post(
f"{base}/v1/systemone",
headers={"Authorization": f"Bearer {key}"},
json={"model": model, "state": state, "questions": questions},
timeout=60,
)
r.raise_for_status()
return r.json()
res = jev(
state={"ticket": "订单被扣两次钱,要求退款", "order": {"charges": [49, 49]}},
questions={
"is_refund": {"type": "noul", "instructions": "用户是否在申请退款"},
"priority": {"type": "score", "instructions": "处理优先级",
"criteria": ["低", "中", "高", "紧急"]},
},
base="https://api.new.bi", key="sk-xxxxxxxx",
)
print(res["answers"]["is_refund"]["noul"])
Node / TypeScript
const BASE = 'https://api.new.bi'
const res = await fetch(`${BASE}/v1/systemone`, {
method: 'POST',
headers: { Authorization: `Bearer ${KEY}`, 'Content-Type': 'application/json' },
body: JSON.stringify({
model: 'jev-1.13.0',
state: '用户反馈支付失败',
questions: { is_bug: { type: 'noul', instructions: '是否产品缺陷' } },
}),
})
const { answers, usage } = await res.json()
三、协议规格
请求
|
字段 |
类型 |
必填 |
说明 |
|---|---|---|---|
|
|
string |
✅ |
固定 |
|
|
string | object | array |
✅ |
待判定上下文,业务数据放这里 |
|
|
object |
✅ |
key 为自定义问题 ID,value 为问题配置 |
state 三种形态:
"state": "用户的原话"
"state": { "message": "订单被扣两次钱", "policy": "重复扣款可退", "order": { "charges": [49, 49] } }
"state": ["第一次反馈:页面空白", "第二次反馈:仍未解决,已影响全天订单"]
对象形态最适合直接丢业务原始数据,不用拼接长文本。
三种问题类型
noul —— 是非判定(输出「是」的概率 0~1)
"is_refund": {
"type": "noul",
"instructions": "用户是否在申请退款",
"criteria": { "true": "明确提出退款诉求", "false": "没有退款请求" }
}
|
字段 |
必填 |
类型 |
|---|---|---|
|
|
✅ |
|
|
|
✅ |
string —— 必须提供 |
|
|
否 |
object |
输出 {"type":"noul","noul":0.97},没有 confidence。
choice —— 单选分类
"team": {
"type": "choice",
"instructions": "归属团队",
"criteria": { "billing": "账单退款", "tech": "功能缺陷", "account": "账号问题" }
}
|
字段 |
必填 |
类型 |
|---|---|---|
|
|
✅ |
|
|
|
✅ |
object,1~255 个选项 |
输出含选中值、全选项概率、置信度。
score —— 有序等级打分
"fury": {
"type": "score",
"instructions": "客户情绪严重程度",
"criteria": ["平静", "沮丧", "非常愤怒"]
}
|
字段 |
必填 |
类型 |
|---|---|---|
|
|
✅ |
|
|
|
✅ |
array,2~10 个等级 |
输出 score 是概率加权值,可落在两个等级之间(如 1.73)。
速查
|
类型 |
|
数量 |
额外必填 |
|---|---|---|---|
|
|
object(可选) |
— |
|
|
|
object |
1~255 |
|
|
|
array |
2~10 |
|
choice 用对象、score 用数组,写反会报错。
响应
{
"model": "jev-1.13.0",
"answers": { "...": {} },
"usage": { "input_tokens": 409, "output_tokens": 69 }
}
三个解析要点:
-
legend是对象{"0":"平静","1":"沮丧"},不是数组 -
score的probabilities也是对象 -
answers的键按字典序重排,与请求顺序无关 —— 一律按问题 ID 取值
四、阈值 + 置信度双闸门
noul / score 给出强度,choice 额外给出置信度。组合起来就是一套可灰度、可回退的决策:
a = res["answers"]
# 概率卡点:高置信才自动处理
if a["is_refund"]["noul"] > 0.85:
auto_refund()
elif a["is_refund"]["noul"] > 0.5:
queue_for_review()
else:
pass
# 分类卡置信度:低了转人工
dept = a["team"]
if dept["confidence"] > 0.7:
assign(dept["choice"])
else:
assign_human("team")
# 等级卡分数
if a["fury"]["score"] >= 1.5:
mark_urgent()
三个原则:
-
阈值可调召回:
noul阈值高则准、低则全 -
置信度决定自动化边界:
confidence低就不要自动执行 -
probabilities别丢:可用于人工队列排序、线上抽检
五、应用场景
1. 电商 / 零售:售后退款判定
{
"model": "jev-1.13.0",
"state": {
"ticket": "订单 A-104 被扣两次钱,请退还重复扣费,已经等了两天没人回",
"policy": "重复扣款属于退款条件",
"order": { "charges": [49, 49], "waited_hours": 48 }
},
"questions": {
"is_refund": { "type": "noul", "instructions": "用户是否在申请退款" },
"reason": {
"type": "choice",
"instructions": "退款原因归类",
"criteria": {
"duplicate": "重复扣款",
"not_received": "未收到货",
"quality": "质量问题",
"change_mind": "无理由退款"
}
},
"urgency": {
"type": "score",
"instructions": "处理紧迫度",
"criteria": ["低", "中", "高", "紧急"]
}
}
}
一次请求完成「是否退款 + 归因 + 优先级」,替代规则引擎里的 if-else 堆叠。reason 直接对接财务系统的退款单类型。
2. 金融 / 风控:交易风险分级
{
"model": "jev-1.13.0",
"state": {
"txn": { "amount": 48000, "currency": "CNY", "channel": "new_device", "hour": 3 },
"user_profile": { "age_days": 2, "avg_amount": 300, "history_chargeback": 0 },
"signals": ["新设备", "凌晨大额", "账户注册仅 2 天"]
},
"questions": {
"is_fraud": { "type": "noul", "instructions": "该交易是否具有欺诈特征" },
"risk_level": {
"type": "score",
"instructions": "风险等级",
"criteria": ["正常", "关注", "可疑", "高危"]
},
"action": {
"type": "choice",
"instructions": "建议处置动作",
"criteria": {
"pass": "放行",
"verify": "二次验证",
"hold": "人工审核",
"block": "拦截"
}
}
}
}
is_fraud 做闸门,risk_level 做分层,action 驱动风控引擎。action 的 confidence 低于阈值时一律降级到 hold,不会误拦。
3. 客服 / SaaS:工单智能路由
{
"model": "jev-1.13.0",
"state": {
"ticket": "企业版账号下三个子账号都无法登录,提示 token 失效,影响今天上线",
"plan": "enterprise",
"sla_hours": 4
},
"questions": {
"team": {
"type": "choice",
"instructions": "工单应分配到的团队",
"criteria": {
"auth": "认证与登录",
"billing": "账单与发票",
"integration": "集成与技术故障",
"success": "客户成功与咨询"
}
},
"impact": {
"type": "score",
"instructions": "业务影响程度",
"criteria": ["无影响", "个别用户", "部分功能", "全线阻断"]
},
"needs_engineer": { "type": "noul", "instructions": "是否需要工程师介入而非客服处理" }
}
}
team + impact 一起决定 SLA 计时与升级路径。企业版 + 全线阻断 → 直接触达值班工程师。
4. 内容平台:审核与创作者治理
{
"model": "jev-1.13.0",
"state": {
"text": "<待审内容原文>",
"author": { "strikes": 2, "account_days": 400 },
"channel": "comment"
},
"questions": {
"is_violation": { "type": "noul", "instructions": "内容是否违反社区规范" },
"category": {
"type": "choice",
"instructions": "违规类目",
"criteria": {
"spam": "垃圾营销",
"harassment": "人身攻击",
"adult": "成人内容",
"illegal": "违法违规",
"normal": "无违规"
}
},
"severity": {
"type": "score",
"instructions": "违规严重程度",
"criteria": ["无", "轻微", "中度", "严重"]
}
}
}
is_violation 卡点做初筛,category 决定处置策略(限流 / 删除 / 封号),severity 结合历史 strikes 决定处罚档位。probabilities 用于人工复审队列排序。
5. 医疗健康:就诊分流
定位说明:Jev 在这里做的是分诊与资源路由,不是诊断,最终判断必须由医护人员完成。
{
"model": "jev-1.13.0",
"state": {
"chief_complaint": "胸口发闷三天,活动后加重,无既往心脏病史",
"vitals": { "age": 52, "bp": "148/92", "smoker": true },
"channel": "online_consult"
},
"questions": {
"urgency": {
"type": "score",
"instructions": "建议就诊紧迫度",
"criteria": ["择期", "尽快", "当日", "急诊"]
},
"department": {
"type": "choice",
"instructions": "建议挂哪个科室",
"criteria": {
"cardiology": "心血管内科",
"respiratory": "呼吸内科",
"gastro": "消化内科",
"general": "全科/内科"
}
},
"needs_offline": { "type": "noul", "instructions": "是否建议线下就诊而非线上咨询" }
}
}
把「急诊 / 当日 / 尽快 / 择期」作为分流档位,配 needs_offline 决定是否打断线上流程。分诊结果必须标记为建议,由医生确认。
6. 教育:学情分层与答疑路由
{
"model": "jev-1.13.0",
"state": {
"student": { "grade": "初二", "recent_scores": [62, 58, 71], "homework_late": 4 },
"submission": "这道题我算出来是负数,但答案说是正数,是不是题目错了",
"subject": "math"
},
"questions": {
"intent": {
"type": "choice",
"instructions": "学生提问的真实意图",
"criteria": {
"answer_check": "对答案有疑问",
"concept": "概念不清",
"method": "解法不会",
"complaint": "抱怨题目/作业量"
}
},
"misconception": { "type": "noul", "instructions": "是否存在明显的概念性错误" },
"engagement": {
"type": "score",
"instructions": "学习投入度",
"criteria": ["消极", "偏低", "正常", "积极"]
}
}
}
intent 决定是答疑还是推练习,misconception 触发针对性微课,engagement 结合迟交次数进干预队列。
7. 物流 / 供应链:异常件归因
{
"model": "jev-1.13.0",
"state": {
"order": { "promised": "48h", "elapsed": "96h", "stops": ["已揽收", "运输中", "运输中"] },
"complaint": "三天了还没到,客服也不回",
"weather_alert": false
},
"questions": {
"cause": {
"type": "choice",
"instructions": "延误原因归类",
"criteria": {
"carrier": "承运商积压",
"address": "地址问题",
"weather": "天气原因",
"ops": "内部操作失误"
}
},
"is_compensable": { "type": "noul", "instructions": "是否属于可赔付情形" },
"churn_risk": {
"type": "score",
"instructions": "客户流失风险",
"criteria": ["低", "中", "高", "极高"]
}
}
}
cause 决定赔付归属(承运商 / 自身),is_compensable 触发赔付流程,churn_risk 决定是否叠加安抚券。
8. 招聘 / HR:简历初筛
{
"model": "jev-1.13.0",
"state": {
"jd": { "title": "后端工程师", "must": ["Go", "分布式", "3 年以上"], "nice": ["K8s"] },
"resume": "<简历文本>"
},
"questions": {
"meets_must": { "type": "noul", "instructions": "是否满足全部硬性要求" },
"match_level": {
"type": "score",
"instructions": "岗位匹配度",
"criteria": ["不匹配", "部分匹配", "较匹配", "高度匹配"]
},
"concern": {
"type": "choice",
"instructions": "最需要面试澄清的点",
"criteria": {
"none": "无",
"gap": "履历空档",
"depth": "技术深度存疑",
"stability": "稳定性存疑"
}
}
}
}
meets_must 做硬性过滤,match_level 排序,concern 直接生成面试提纲。保留人工复核环节,避免把筛选完全交给模型。
9. Agent / 工程:工具调用护栏
{
"model": "jev-1.13.0",
"state": {
"user_request": "把上个月的报表全部删掉",
"pending_tool_call": { "name": "delete_reports", "args": { "period": "last_month", "scope": "all" } },
"policy": "删除类工具需用户显式确认,且不允许批量删除"
},
"questions": {
"should_execute": { "type": "noul", "instructions": "该工具调用是否应直接执行" },
"risk": {
"type": "score",
"instructions": "操作风险等级",
"criteria": ["安全", "低风险", "高风险", "危险"]
},
"violated_rule": {
"type": "choice",
"instructions": "违反的策略条目",
"criteria": {
"none": "未违反",
"no_confirm": "缺少显式确认",
"bulk": "批量操作限制",
"scope": "越权范围"
}
}
}
}
Agent 执行工具前先过一遍,should_execute 为假则中断并回问用户,violated_rule 直接生成解释文案。比让模型「自己想一想」稳定且可量化。
10. 保险:理赔单分流
{
"model": "jev-1.13.0",
"state": {
"claim": { "type": "医疗险", "amount": 86000, "incident_days_ago": 3 },
"docs": ["诊断证明", "发票"],
"missing": ["出院小结"]
},
"questions": {
"is_complete": { "type": "noul", "instructions": "材料是否齐全" },
"lane": {
"type": "choice",
"instructions": "应进入哪条处理通道",
"criteria": {
"auto": "自动快速理赔",
"normal": "常规人工审核",
"investigate": "需调查",
"reject_risk": "疑似不实申报"
}
},
"fraud_probability": {
"type": "score",
"instructions": "疑似欺诈程度",
"criteria": ["无", "低", "中", "高"]
}
}
}
is_complete 决定是否先补件,lane 决定人力投入,fraud_probability 触发调查流程。
六、工程实践
一次问完,别串行
最常见的误用是一次只问一个问题,连发三次请求。Jev 的设计就是一次问多组、并行独立判定,串行既慢又浪费调用次数。
问题之间不可见
同一个请求里的 questions 互相看不到对方的答案。所以不能做「根据上一题结果决定下一题」的链式推理。需要链式,就分两次请求。
用对象形态传业务数据
# 推荐:原样丢结构化数据
state = {"order": {...}, "user": {...}, "policy": "..."}
而不是自己拼成一段长文本 —— 对象形态信息损失更小。
全部走置信度兜底
def decide(answer, auto_threshold=0.7):
if answer["type"] == "choice" and answer["confidence"] < auto_threshold:
return {"action": "human_review", "suggest": answer["choice"]}
return {"action": "auto", "value": answer.get("choice") or answer.get("noul")}
错误处理
try:
res = jev(state, questions, base=BASE, key=KEY)
except requests.HTTPError as e:
if e.response.status_code == 503:
... # 网关无可用渠道,检查渠道配置
elif e.response.status_code in (429, 500, 502, 503):
... # 重试
raise
日志留痕
把 usage.input_tokens 和 answers 一起记下来 —— Jev 只收输入 token 费用,input_tokens 就是成本,便于按业务线核算。
七、限制与排错
限制
|
项 |
限制 |
|---|---|
|
上下文总长 |
64k token |
|
|
32k token |
|
输入模态 |
仅文本,不支持图片 / 音频 / 视频 |
|
语言 |
英文训练为主,中文可用但建议自测精度 |
|
|
2~10 |
|
|
1~255 |
无效参数
temperature、max_tokens、stream、response_format 对 Jev 无意义。stream: true 不会返回 SSE,始终返回单个 JSON。
报错速查
|
报错 |
原因 |
|---|---|
|
|
缺 |
|
|
缺 |
|
|
缺 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
单条 |
|
|
缺 |
|
HTTP 503 |
网关路由未覆盖该路径 |
八、落地清单
-
一次请求问多组问题 —— 不要串行
-
state用对象形态 —— 直接丢业务原始数据 -
noul必写instructions—— 只写criteria会报错 -
choice用对象、score用数组 —— 类型别写反 -
阈值 + 置信度双闸门 —— 高强度才自动,低置信转人工
-
按问题 ID 取值 ——
answers键序是字典序 -
legend/probabilities按对象解析 —— 不是数组 -
保留人工兜底 —— 尤其医疗、金融、招聘等强合规场景
312



被折叠的 条评论
为什么被折叠?



