Jev 案例：简历技能标签抽取（完整代码）

TL;DR： 本案例为招聘平台搭建一条简历打标流水线：第一次是多选题，把自由文本简历映射到一张固定的 12 个标准化技能标签表上；第二次是打分题，给每个提取出的技能打一个 0–100 的熟练度置信分。完整请求体、示例数据（example fixture）格式的响应、可直接运行的 Python 脚本，以及与正则/关键词匹配的对比讨论都在下文。

场景

我们给一家招聘平台做打标后端，每周大约进 4000 份简历。每份简历都要打上技能标签供招聘方筛选，而且标签必须来自一张固定词表——招聘人员搜索的是 python，不是”有 Python 数据管道经验”，标签和词表差一个词，所有筛选器里它就等于不存在。

这条流水线之前，打标靠人工。两个问题叠加在一起：成本和一致性。成本在每周 4000 份的量上很好算。一致性是更麻烦的问题——同一份简历，一个外包会把”built AWS Lambda functions”标成 aws，另一个会直接跳过，因为简历里没出现光杆的”AWS”两个字。同一份简历不同人打出的标签不一样，下游所有筛选器的数据就这样悄悄被污染了。

我们把人工打标换成了每份简历两次 Jev 调用：

1. 多选题——从固定的 12 个标签（python、sql、aws、docker、communication 等共 12 个）里，把简历有证据支撑的标签全部选出来。这里必须用多选：单选题会逼出一个失真的”主技能”，而招聘人员会过度看重这个标签。
2. 打分题——对每个提取出的技能，按 0–100 打一个分数，衡量简历对”真实熟练度”（而不是随口一提）的支持程度。

Jev 是 TypeSafe AI 在 2026 年 9 月发布的 System One 判断模型。它的类型化 JSON 输出每个答案都带置信度，这才让第二次调用有意义——一个不带置信度的分数只会把一致性问题往下游推。下文的示例简历刻意用了同义词和项目描述，而不是规整的技能清单，这恰恰是关键词匹配最吃亏的地方。

请求

第一次调用——标签提取（多选题）：

{
  "model": "typesafe/jev-1.13",
  "messages": [
    {
      "role": "system",
      "content": "Multi-select choice question. From this fixed tag list: python, sql, aws, docker, kubernetes, react, machine-learning, data-analysis, project-management, communication, leadership, customer-service — select every tag the resume clearly supports. Do not infer tags from job titles alone; require concrete evidence in the text. Reply with JSON only: {\"answer\": [<tag>, ...], \"confidence\": <0-1>, \"rationale\": <one sentence>}"
    },
    {
      "role": "user",
      "content": "Resume: Data analyst with 4 years of experience. Built ETL pipelines in Python pulling from a Postgres warehouse, wrote daily reporting queries in SQL. Migrated our team's batch jobs to AWS (Lambda + S3) and containerized local dev with Docker. Led a 3-person analytics pod, presented findings to non-technical stakeholders every sprint. Side project: a small scikit-learn model predicting churn."
    }
  ]
}

第二次调用——熟练度打分（打分题），每个提取出的技能一次调用：

{
  "model": "typesafe/jev-1.13",
  "messages": [
    {
      "role": "system",
      "content": "Scoring question. Rate 0-100 how confidently this resume demonstrates working proficiency with the given skill — 0 means no evidence, 50 means mentioned without depth, 100 means sustained professional use with concrete outcomes. Reply with JSON only: {\"answer\": <0-100>, \"confidence\": <0-1>, \"rationale\": <one sentence>}"
    },
    {
      "role": "user",
      "content": "Skill: aws. Resume excerpt: \"Migrated our team's batch jobs to AWS (Lambda + S3).\""
    }
  ]
}

响应

以下两个响应均为示例数据（example fixture），用于说明格式，不是线上实测抓包。

第一次调用：

{
  "answer": ["python", "sql", "aws", "docker", "leadership", "communication", "machine-learning"],
  "confidence": 0.86,
  "rationale": "Each tag has concrete textual evidence: pipelines in Python, SQL reporting queries, Lambda/S3 migration, Docker for dev, a led pod, stakeholder presentations, and a scikit-learn project."
}

第二次调用（针对 aws）：

{
  "answer": 65,
  "confidence": 0.83,
  "rationale": "A real production migration is described, but scope is one team's batch jobs with no scale or outcome details."
}

复现脚本

保存为 skill_tags.py，填入你自己的 BASE_URL 和 API key 后运行 python3 skill_tags.py。脚本使用 OpenAI 兼容 SDK，任何 OpenAI 兼容端点都能跑：

import json, os
from openai import OpenAI

BASE_URL = "https://YOUR-ENDPOINT.example.com/v1"  # replace with your provider
client = OpenAI(base_url=BASE_URL, api_key=os.environ["JEV_API_KEY"])
MODEL = "typesafe/jev-1.13"  # confirm exact slug with your provider

TAGS = ("python, sql, aws, docker, kubernetes, react, machine-learning, "
        "data-analysis, project-management, communication, leadership, "
        "customer-service")

def ask(system, user):
    res = client.chat.completions.create(
        model=MODEL,
        messages=[{"role": "system", "content": system},
                  {"role": "user", "content": user}])
    return json.loads(res.choices[0].message.content)

EXTRACT = ('Multi-select choice question. From this fixed tag list: ' + TAGS +
           ' — select every tag the resume clearly supports. Do not infer '
           'tags from job titles alone; require concrete evidence in the '
           'text. Reply with JSON only: {"answer": [<tag>, ...], '
           '"confidence": <0-1>, "rationale": <one sentence>}')

SCORE = ('Scoring question. Rate 0-100 how confidently this resume '
         'demonstrates working proficiency with the given skill — 0 means '
         'no evidence, 50 means mentioned without depth, 100 means '
         'sustained professional use with concrete outcomes. Reply with '
         'JSON only: {"answer": <0-100>, "confidence": <0-1>, '
         '"rationale": <one sentence>}')

resume = ("Data analyst with 4 years of experience. Built ETL pipelines in "
          "Python pulling from a Postgres warehouse, wrote daily reporting "
          "queries in SQL. Migrated our team's batch jobs to AWS (Lambda + "
          "S3) and containerized local dev with Docker. Led a 3-person "
          "analytics pod, presented findings to non-technical stakeholders "
          "every sprint. Side project: a small scikit-learn model "
          "predicting churn.")

tags = ask(EXTRACT, "Resume: " + resume)
print(tags)

tagged = {}
for tag in tags["answer"]:
    scored = ask(SCORE, "Skill: " + tag + ". Resume excerpt: " + resume)
    if scored["answer"] >= 40 and scored["confidence"] >= 0.5:
        tagged[tag] = scored["answer"]
print(json.dumps(tagged, indent=2))

结尾这道后置过滤（分数 ≥ 40 且置信度 ≥ 0.5）正是 12 标签固定词表发挥作用的地方：把弱证据丢掉，标签集才够小，招聘人员才会信它。

为什么不用正则或关键词匹配？

关键词匹配仍然值得保留作为基线：它免费、即时，能兜住那些字面命中的场景，否则这些场景也要花一次模型调用。我们的做法是先跑关键词，只把关键词阶段搞不定的简历（或标签候选）送进 Jev。这条流水线不是替代便宜方案，而是接住便宜方案漏掉的部分。

关键参数

注意事项

• 上面的响应均为示例数据（example fixture）；字段名请与官方文档核对后再依赖。官方 API 细节在 typesafe.ai，入口是System One 模型发布公告。
• 原始 HTTP 响应是 OpenAI 兼容的 chat completion 结构；本文展示的 JSON 是从 message content 里解析出来的模型类型化输出。
• 每个提取出的标签都要一次打分调用，标签多的简历成本会翻着涨。预算紧张时，按提取置信度取前 8 个标签再送打分。
• 改动标签词表后，先对一份固定的旧简历样本重新打标并 diff 前后差异，再上线——改词表会连带改变旧标签。
• 适用于 Jev 1.13（2026 年 9 月发布），演示渠道为 OpenRouter。