issue-review.yml

name: Issue Review

on:
  issues:
    types: [opened]
permissions:
  issues: write
  contents: read
jobs:
  review_issue:
    runs-on: ubuntu-latest
    
    steps:
    - name: Checkout repository
      uses: actions/checkout@v2

    - name: Set up Python
      uses: actions/setup-python@v2
      with:
        python-version: '3.x'

    - name: Install dependencies
      run: |
        python -m pip install --upgrade pip
        pip install openai
        pip install PyGithub
        pip install qdrant-client
        pip install regex

    - name: Review Issue with LLM
      env:
        OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        QD_API_KEY: ${{ secrets.QD_API_KEY }}
        QD_URL: ${{ secrets.QD_URL }}
      run: |
        python <<EOF
        import github
        from github import Github
        import os
        import openai
        import regex as re
        from qdrant_client import QdrantClient
        from qdrant_client.models import PointStruct

        token = os.getenv("GITHUB_TOKEN")
        qd_api = os.getenv("QD_API_KEY")
        qd_url = os.getenv("QD_URL")
        g = Github(token)
        repo = g.get_repo("${{ github.repository }}")
        issue = repo.get_issue(${{ github.event.issue.number }})
        issue_content = f"{issue.title}\n{issue.body}"
        try:
            repo.create_label(name="toxic", color="ff0000")
            repo.create_label(name="duplicated", color="708090")
        except:
            pass

        qdrant_client = QdrantClient(
            url=qd_url, 
            api_key=qd_api,
        )
        openai_client = openai.Client()
        embedding_model = "text-embedding-3-small"
        collection_name = "issue_collection"


        def validate_image(text):
            model_name = "gpt-4o"
            prompt = "この画像が暴力的、もしくは性的な画像の場合trueと返してください。"
            image_url = re.search(r"!\[[^\s]+\]\((https://[^\s]+)\)", text)
            if image_url and len(image_url) > 1:
                image_url = image_url[1]
            else:
                return False
            try:
                response = openai_client.chat.completions.create(
                  model=model_name,
                  messages=[
                    {
                      "role": "user",
                      "content": [
                        {"type": "text", "text": prompt},
                        {
                          "type": "image_url",
                          "image_url": {
                            "url": image_url
                          },
                        },
                      ],
                    }
                  ],
                  max_tokens=1200,
                )
            except:
                return True
            v = response.choices[0].message.content.lower()
            if "true" in v:
                return True
            else:
                return False

        def judge_violation(text):
            response = openai_client.moderations.create(input=text)
            flag = response.results[0].flagged
            video_flag = validate_image(text)
            if flag or video_flag:
                print(response)
                issue.add_to_labels("toxic")
                if video_flag
                    warn = "不適切な画像です。アカウントBANの危険性があります。"
                else:
                    warn = "不適切な投稿です。アカウントBANの危険性があります。"
                issue.create_comment(warn)
                issue.edit(state="closed")
                return True
            return flag

        def add_issue(text:str, iss_num:int):
            texts = [text]
            ids = [iss_num]
            result = openai_client.embeddings.create(input=texts, model=embedding_model)
            points = [
                PointStruct(
                    id=idx,
                    vector=data.embedding,
                    payload={"text": t},
                )
                for idx, data, t in zip(ids, result.data, texts)
            ]
            qdrant_client.upsert(collection_name, points)
            return text

        def merge_issue(iss:int):
            issue.add_to_labels("duplicated")
            print(f"merge to {iss}")
            issue.create_comment(f"#{iss} と重複しているかもしれません")
            return iss

        def qd_search(text:str):
            results = qdrant_client.search(
                collection_name=collection_name,
                query_vector=openai_client.embeddings.create(
                    input=[text],
                    model=embedding_model,
                )
                .data[0]
                .embedding,
            )
            return results

        def qd_add(text:str, iss_num:int):
            texts = [text]
            ids = [iss_num]
            result = openai_client.embeddings.create(input=texts, model=embedding_model)
            points = [
            PointStruct(
                id=idx,
                vector=data.embedding,
                payload={"text": text},
            )
            for idx, data, text in zip(ids, result.data, texts)
            ]
            qdrant_client.upsert(collection_name, points)


        if judge_violation(issue_content):
            quit()
        results = qd_search(issue_content)

        if len(results) > 2:
            results = results[:3]
        else:
            results = results
        print(results)
        res = ""
        for i in results:
            res+=f'id:{i.id}\n内容:{i.payload["text"]}\n'
        res = res.strip()

        prompt= f"""
        以下は市民から寄せられた政策提案です。
        {issue_content}
        この投稿を読み、以下の過去提案の中に重複する提案があるかを判断してください。
        {res}
        重複する提案があればそのidを出力してください。
        もし存在しない場合は0と出力してください。

        [出力形式]
        id:0
        """
        print(prompt)
        completion = openai_client.chat.completions.create(
          model="gpt-4o",
          max_tokens= 1024,
          messages=[
          {"role": "system", "content": prompt},
          ]
        )
        review = completion.choices[0].message.content
        if ":" in review:
            review = review.split(":")[-1]
        if review.isdecimal():
            if review == "0":
                add_issue(issue_content, issue.number)
            else:
                merge_issue(int(review))
        print(review)

        EOF