<div id="readability-page-1" class="page"><div>
                <div id="SE-9c2ed588-fa79-4cb6-87b0-d5713f10e0d5">
                                    <p id="SE-4abc1c0f-8ce0-42c2-813f-e2e415d98a1a"><span id="SE-fb1196ca-9e23-485b-ba7e-0cf6f4ba631b">안녕하새요 멍개입니다.</span></p>
<p id="SE-78a65149-6a1a-48ff-b343-52bd305559bf"><span id="SE-42543775-b051-4303-b98a-e5a128a2c88a">​</span></p>
<p id="SE-4b5964ae-99ba-4a61-8bd2-8106fef3b00a"><span id="SE-beb78ed1-81b5-4f24-abb3-c2bb1f5f2626">지난시간에 TypeSafe에서 공개한 jev를 알아보고 Qwen2.5 1 B RLCD으로 jev를 흉내낸 프로젝트를 구동시켜보았습니다.</span></p>
<p id="SE-05b09842-e2ae-4d21-a24c-af4e40371b36"><span id="SE-c22814cd-d677-4cd7-80fa-d7315af69081">​</span></p>
<p id="SE-751d87ea-4442-477e-845a-6568ca77ef99"><span id="SE-faa4a27f-0b7f-43be-8b9b-baad3e021253">기능적으로는 같지만 요청/응답 스키마가 조금 달라 typesafe sdk를 사용할 수 없다는 단점이 있어 jev와 동일한 요청/응답 스키마를 구현한 오픈소스를 찾던도중 open jev라는 것을 발견하게 되었습니다.</span></p>
<p id="SE-054ae026-d603-4510-bc0c-834992751400"><span id="SE-d7daebf8-3ea3-4bb2-a2c5-859cc9799df4"><a href="https://blog.naver.com/pjt3591oo/224416016792" data-linktype="text" data-linkdata="{&quot;id&quot;: &quot;SE-d7daebf8-3ea3-4bb2-a2c5-859cc9799df4&quot;, &quot;link&quot;: &quot;https://blog.naver.com/pjt3591oo/224416016792&quot;}" target="_blank">https://blog.naver.com/pjt3591oo/224416016792</a></span></p>

                            </div>                
                <div id="SE-ec5ab84e-45c5-4a04-8a01-887d879aa2af">
                            <p id="SE-d4078151-1172-40b7-b7d2-412ba5bc9a2f"><span id="SE-a111659c-f786-43a6-ae5b-72e18775047e">참고로 TypeSafe 웨이트 리스트는 하루 정도면 통과됩니다. 저도 어제 요청했는데 하루만에 승인되었습니다.</span></p>
                        </div>                <div id="SE-380423c7-a689-41db-8b8e-23fc8cef4a25">
                                <p><a href="#" onclick="return false;" data-linktype="img" data-linkdata="{&quot;id&quot; : &quot;SE-380423c7-a689-41db-8b8e-23fc8cef4a25&quot;, &quot;src&quot; : &quot;https://mblogthumb-phinf.pstatic.net/MjAyNjA5MTlfMTMy/MDAxNzg5Nzk0NjYyMzUz.1maQ2Owrit_s6o0rd8BV8gBEOCQGOM9gAgPuaQGmc5wg.-SepANabFCvak-QFUe6Zh9Dx47kKn-lIUuOXpA-yApUg.PNG/%EC%8A%A4%ED%81%AC%EB%A6%B0%EC%83%B7_2026-09-19_%EC%98%A4%ED%9B%84_2.10.54.png&quot;, &quot;originalWidth&quot; : &quot;1267&quot;, &quot;originalHeight&quot; : &quot;1115&quot;, &quot;linkUse&quot; : &quot;false&quot;, &quot;link&quot; : &quot;&quot;, &quot;fileSize&quot;: &quot;170879&quot; }">
                                        <img src="https://mblogthumb-phinf.pstatic.net/MjAyNjA5MTlfMTMy/MDAxNzg5Nzk0NjYyMzUz.1maQ2Owrit_s6o0rd8BV8gBEOCQGOM9gAgPuaQGmc5wg.-SepANabFCvak-QFUe6Zh9Dx47kKn-lIUuOXpA-yApUg.PNG/%EC%8A%A4%ED%81%AC%EB%A6%B0%EC%83%B7_2026-09-19_%EC%98%A4%ED%9B%84_2.10.54.png?type=w800" data-width="693" data-height="609" alt="">
                                    </a>
                                </p>
                        </div>
                                <div id="SE-fead865c-5513-4f34-9cfb-62de03a735b3">
                                <p><a href="#" onclick="return false;" data-linktype="img" data-linkdata="{&quot;id&quot; : &quot;SE-fead865c-5513-4f34-9cfb-62de03a735b3&quot;, &quot;src&quot; : &quot;https://mblogthumb-phinf.pstatic.net/MjAyNjA5MTlfNSAg/MDAxNzg5Nzk0Njc0MDc4.cFXe9o7Bj2LR7IAai0gB8BtoeS5RuNq5i7LjwVpoSGYg.vpdwCZKwBvVNdStBhdx6IMiPLUlIL4wLUKd1pzRjxFkg.PNG/%EC%8A%A4%ED%81%AC%EB%A6%B0%EC%83%B7_2026-09-19_%EC%98%A4%ED%9B%84_2.11.07.png&quot;, &quot;originalWidth&quot; : &quot;1267&quot;, &quot;originalHeight&quot; : &quot;1115&quot;, &quot;linkUse&quot; : &quot;false&quot;, &quot;link&quot; : &quot;&quot;, &quot;fileSize&quot;: &quot;156004&quot; }">
                                        <img src="https://mblogthumb-phinf.pstatic.net/MjAyNjA5MTlfNSAg/MDAxNzg5Nzk0Njc0MDc4.cFXe9o7Bj2LR7IAai0gB8BtoeS5RuNq5i7LjwVpoSGYg.vpdwCZKwBvVNdStBhdx6IMiPLUlIL4wLUKd1pzRjxFkg.PNG/%EC%8A%A4%ED%81%AC%EB%A6%B0%EC%83%B7_2026-09-19_%EC%98%A4%ED%9B%84_2.11.07.png?type=w800" data-width="693" data-height="609" alt="">
                                    </a>
                                </p>
                        </div>
                                <div id="SE-ca06be75-fca0-48e1-bd81-40b57218e073">
                                    <p>$ https://github.com/daseinlabs/open-jev
$ cd open-jev</p>
                                </div>                <div id="SE-c5693c36-ba9f-414b-8a8c-e8faf0b7459a">
                                    <p id="SE-a68101ab-a0d8-427d-8e84-8e061a396ee6"><span id="SE-1c4f22be-d9da-4166-a075-45f4e12e2cd1">해당 프로젝트는 make를 기반으로 프로젝트를 구성하고 실행하도록 설정되어 있습니다.</span></p>
<p id="SE-dc58b84f-6261-4f6d-804e-fa5a51a19e7b"><span id="SE-9e7279f7-08f8-4261-860b-3f3a017388b2">​</span></p>
<p id="SE-dabd5235-0a3d-4f31-bc3d-f1575c92eb81"><span id="SE-371fcafa-b3b8-4829-902d-4229d05744ed">Makefile을 열어보면 여러 설정들을 확인할 수 있습니다.  open-jev는 기본적으로 google/gemma-3-4b-it 모델을 사용하고 있는데 여기서는 qwen2.5-1.5b-instruct-4bit 모델을 활용해보겠습니다. 머 이건 각자 원하는 모델을 사용하면 됩니다.</span></p>
<p id="SE-5c3cc629-f56d-402d-bd30-2a79c2925a36"><span id="SE-8541efb2-82b1-4b99-b4b8-c683f54f9e19">​</span></p>
<p id="SE-72a834b4-21a9-498a-bec6-0fae86427c2c"><span id="SE-c3a6bf6d-dbc5-455e-8164-8b3a7282e857">make setup을 수행하면 의존성을 설치하고 모델을 받습니다. </span></p>

                            </div>                <div id="SE-c7355415-bf13-40ab-8fed-f202f033f632">
                                    <p>$ export MODEL=models/Qwen2.5-1.5B-Instruct-4bit
$ export HF_REPO=mlx-community/Qwen2.5-1.5B-Instruct-4bit

$ make setup</p>
                                </div>                                <div id="SE-60b2d626-8be6-4e4c-b697-6f639ee1933d">
                                    <p>$ make serve

.venv/bin/openjev serve --host 127.0.0.1 --port 8000 --model models/Qwen2.5-1.5B-Instruct-4bit
INFO:     Started server process [35606]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://127.0.0.1:8000 (Press CTRL+C to quit)</p>
                                </div>                <div id="SE-9a675d52-cf1b-4bc1-aa23-2c2fb99d5c36">
                                    <p id="SE-75d14f04-bc6f-40fc-b5e1-c504f9043a7f"><span id="SE-6dd2c377-9c1a-44e4-8d45-c78bf603bc49">준비는 완료되었습니다.</span></p>
<p id="SE-ba91324d-fa14-4e03-954f-f476f8fff678"><span id="SE-1cd91120-3e39-4ef0-84d2-ebec77af0b8a">​</span></p>
<p id="SE-83bb1c7a-95d8-43ca-ad77-00625e801601"><span id="SE-a2155282-de7c-4dbb-af84-e456c1f75f0f"><b>· jev 사용</b></span></p>
<p id="SE-46cf8204-5302-4e5e-a73d-b2e584fe128d"><span id="SE-e84e6e07-de82-4531-93ad-8f6772329319">먼저, http를 직접 요청해보겠습니다.</span></p>
<p id="SE-d39ccc54-915d-4c31-b929-a7bd6ad8af18"><span id="SE-2908c515-19d9-4332-bc68-710b94ce51dc">​</span></p>
<p id="SE-91b81d90-772c-415c-915e-2b92b1c73acb"><span id="SE-ca922c1a-4b56-4e64-8265-3a7d5e73d609">jev는 POST /v1/systemone를 사용합니다.</span></p>

                            </div>                <div id="SE-8ea4f66b-2895-43de-9509-4e497dacdd82">
                                    <p>$ curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "I ordered size 10 shoes but received size 8. Please send the right size.",
  "model": "jev-latest",
  "questions": {
    "department":  {"type": "choice", "instructions": "Which department handles this?",
                    "criteria": {"returns": "Returns and exchanges", "shipping": "Delivery issues", "billing": "Charges and refunds"}},
    "severity":    {"type": "score",  "instructions": "How severe is the problem?",
                    "criteria": ["minor", "moderate: wrong item", "major: safety or financial loss"]},
    "wants_refund":{"type": "noul",   "instructions": "Is the customer asking for a refund to their card?"}
  }
}'</p>
                                </div>                <div id="SE-3af31fab-959c-429c-afa6-555d5b6b4574">
                                    <p>{
  "model": "jev-latest",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "returns",
      "probabilities": {
        "returns": 0.830675189418666,
        "shipping": 0.0005629096469691894,
        "billing": 0.16876190093436477
      },
      "confidence": 0.5825750418639104
    },
    "severity": {
      "type": "score",
      "score": 1.0232376733773803,
      "confidence": 0.89836225702385,
      "legend": {
        "0": "minor",
        "1": "moderate: wrong item",
        "2": "major: safety or financial loss"
      },
      "probabilities": {
        "0": 0.00009097296195690015,
        "1": 0.9765803806987058,
        "2": 0.02332864633933725
      }
    },
    "wants_refund": {
      "type": "noul",
      "noul": 0.011687257995694434
    }
  },
  "usage": {
    "input_tokens": 165,
    "output_tokens": 8
  }
}</p>
                                </div>                <div id="SE-f520323b-7c1a-4a1c-b182-e01bc4478d67">
                            <p id="SE-fa411d9e-d99c-495d-8f5e-7182eb248bd3"><span id="SE-d28ef8c2-b458-428c-90b6-ea349c89e5a7">만약 SDK를 이용한다면 다음과 같이 환경변수만 바꿔주면 됩니다.</span></p>
                        </div>                <div id="SE-11ae30dd-8ba7-41da-afc0-a95603ced356">
                                    <p>export TYPESAFE_BASE_URL=http://127.0.0.1:8000
export TYPESAFE_API_KEY=something_input</p>
                                </div>                <div id="SE-c956716c-bd62-4fd5-af9b-3f80297b2a47">
                                    <p>from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

for i in range(10):

  with TypeSafeClient() as client:
      response = client.system_one(
          state={"document": "I was charged twice. Please fix this ASAP."},
          questions={
              "billing": Noul(instructions="Is this ticket about billing?"),
              "tone": Choice(
                  instructions="What is the customer's tone?",
                  criteria={"calm": None, "frustrated": None, "angry": None},
              ),
              "urgency": Score(
                  instructions="How urgent is this ticket?",
                  criteria=["can wait", "this week", "today"],
              ),
          },
      )

  print(response)</p>
                                </div>                <div id="SE-8d89bcee-d394-4715-be93-6e01e47e29b4">
                                <p><a href="#" onclick="return false;" data-linktype="img" data-linkdata="{&quot;id&quot; : &quot;SE-8d89bcee-d394-4715-be93-6e01e47e29b4&quot;, &quot;src&quot; : &quot;https://mblogthumb-phinf.pstatic.net/MjAyNjA5MTlfMjA1/MDAxNzg5Nzk1MzM0MzQ4.pzqjT12wBHOe84GPj4JX74BpDUgvQcAJRPt7o1w30aAg.h2_0P7xrLQ_N7MAnYwI_yu08EhPvYL6calQdeRTvFWog.PNG/%EC%8A%A4%ED%81%AC%EB%A6%B0%EC%83%B7_2026-09-19_%EC%98%A4%ED%9B%84_2.22.09.png&quot;, &quot;originalWidth&quot; : &quot;1618&quot;, &quot;originalHeight&quot; : &quot;687&quot;, &quot;linkUse&quot; : &quot;false&quot;, &quot;link&quot; : &quot;&quot;, &quot;fileSize&quot;: &quot;1910930&quot; }">
                                        <img src="https://mblogthumb-phinf.pstatic.net/MjAyNjA5MTlfMjA1/MDAxNzg5Nzk1MzM0MzQ4.pzqjT12wBHOe84GPj4JX74BpDUgvQcAJRPt7o1w30aAg.h2_0P7xrLQ_N7MAnYwI_yu08EhPvYL6calQdeRTvFWog.PNG/%EC%8A%A4%ED%81%AC%EB%A6%B0%EC%83%B7_2026-09-19_%EC%98%A4%ED%9B%84_2.22.09.png?type=w800" data-width="693" data-height="294" alt="">
                                    </a>
                                </p>
                        </div>
                <div id="SE-be7e734c-96cf-4a61-a9a2-552d32b391c5">
                            <p id="SE-3b595d8e-9553-4fe0-9444-4960223f789c"><span id="SE-bfbb6ccb-2932-4d41-bbe6-b7ee68a8b9cf">또한 langchain에서도 typesafe를 사용할 수 있습니다.</span></p>
                        </div>                <div id="SE-458d424e-f14f-4347-886d-cf14d8338783">
                                    <p>$ uv add langchain-typesafe</p>
                                </div>                <div id="SE-54de8167-36ec-42a9-b154-4342b0decddd">
                                    <p>from langchain_typesafe import Choice, Noul, Score, TypeSafeClassifier

classifier = TypeSafeClassifier(
    questions={
        "billing": Noul(
            instructions="Is this ticket about billing?"
        ),
        "tone": Choice(
            instructions="What is the customer's tone?",
            criteria={
                "calm": None,
                "frustrated": None,
                "angry": None,
            },
        ),
        "urgency": Score(
            instructions="How urgent is this ticket?",
            criteria=[
                "can wait",
                "this week",
                "today",
            ],
        ),
    }
)

response = classifier.invoke(
    "I was charged twice. Please fix this ASAP."
)

print(response)</p>
                                </div>                <div id="SE-8f7aac16-f99e-4ee3-8f84-3271a9f933d2">
                            <p id="SE-79f4510e-5658-412e-a2b7-7d00971476ed"><span id="SE-c1eaa0f5-a728-453a-b125-6f51fc7b3344">또한 랭체인에서는 모델 라우팅 기능도 사용할 수 있습니다.</span></p>
                        </div>                <div id="SE-a4673165-6b54-41ab-8a37-5cd478735c14">
                                    <p>from langchain_typesafe.experimental.middleware import (
    ModelChoice,
    ModelRouterMiddleware,
)
from langchain.agents import create_agent

router = ModelRouterMiddleware(
    choices={
        "fast": ModelChoice(
            model="openai:luna",
            criteria="단순 조회, 추출, 간단한 수정",
        ),
        "powerful": ModelChoice(
            model="openai:sol",
            criteria="복잡한 분석, 설계, 어려운 추론",
        ),
    },
    instructions="요청을 처리할 수 있는 적절한 모델을 선택하세요.",
)


agent = create_agent(
    "openai:gpt-5.6-luna",
    middleware=[router],
)</p>
                                </div>                    </div></div>

삽질하는 멍개씨

안녕하새요 멍개입니다. 지난시간에 TypeSafe에서 공개한 jev를 알아보고 Qwen2.5 1 B RLCD으로 jev를 흉내낸 프로젝트를 구동시켜보았습니다. 기능적으로는 같지만 요청/응답 스키마가 조금 달라 typesafe sdk를 사용할 수 없다는 단점이 있어 jev와 동일한 요청/응답 스키마를 구현한 오픈소스를 찾던도중 open jev라는 것을 발견하게 되었습니다. https://blog.naver.com/pjt3591oo/224416016792 참고로 TypeSafe 웨이트 리스트는 하루 정도면 통과됩니다. 저도 어제 요청했는데 하루만에 승인되었습니다. 처음 이용하면 크레딧 $5를 줍니다 · 환경구성 해당 프로젝트는 make를 기반으로 프로젝트를 구성하고 실행하도록 설정되어 있습니다. M.......

로컬에서 jev를 돌려보자

댓글