name: Examples - RAG
permissions:
  contents: read
on:
  schedule:
    # Every day at 6 AM UTC+8
    - cron: '0 22 * * *'

  workflow_dispatch:

  repository_dispatch:
    types: [ci-rag, ci-all]

run-name: >-
  ${{ github.event_name == 'repository_dispatch'
      && format(
        'RAG - PR #{0} - {1} - {2}',
        github.event.client_payload.pull_number,
        github.event.client_payload.ci_label,
        github.event.client_payload.correlation_id
      )
      || format('RAG - {0}', github.event_name) }}

jobs:
  rag:
    if: >
      github.event_name != 'repository_dispatch' ||
      github.event.action == 'ci-rag' ||
      github.event.action == 'ci-all'
    name: RAG (Python ${{ matrix.python-version }}, ${{ matrix.setup-script }})
    runs-on: [self-hosted, 1ES.Pool=agl-runner-gpu]
    timeout-minutes: 60
    strategy:
      matrix:
        include:
        - python-version: '3.10'
          setup-script: 'legacy'
        - python-version: '3.12'
          setup-script: 'stable'
        - python-version: '3.13'
          setup-script: 'latest'
      fail-fast: false
    steps:
      - name: Check GPU status
        run: nvidia-smi
      - name: Check disk space
        run: df -h
      - uses: actions/checkout@v6
        with:
          ref: ${{ github.event_name == 'repository_dispatch' && github.event.client_payload.pr_ref || (github.event.pull_request.number && format('refs/pull/{0}/merge', github.event.pull_request.number)) || github.ref }}
      - uses: astral-sh/setup-uv@v7
        with:
          enable-cache: true
          python-version: ${{ matrix.python-version }}
      - name: Upgrade dependencies (latest)
        run: uv lock --upgrade
        if: matrix.setup-script == 'latest'
      - name: Sync dependencies (latest)
        run: |
          uv sync --frozen --no-default-groups --extra verl \
            --group dev --group experiment --group agents --group rag --group torch-gpu-stable
        if: matrix.setup-script == 'latest'
      - name: Sync dependencies (stable & legacy)
        run: |
          uv sync --frozen --no-default-groups --extra verl \
            --group dev --group experiment --group agents --group rag --group torch-gpu-${{ matrix.setup-script }}
        if: matrix.setup-script != 'latest'
      - name: Freeze dependencies
        run: |
          set -ex
          uv pip freeze | tee requirements-freeze.txt
          echo "UV_LOCKED=1" >> $GITHUB_ENV
          echo "UV_NO_SYNC=1" >> $GITHUB_ENV
      - name: Upload dependencies artifact
        uses: actions/upload-artifact@v6
        with:
          name: dependencies-rag-${{ matrix.python-version }}-${{ matrix.setup-script }}
          path: requirements-freeze.txt
          compression-level: 0

      - name: Launch LiteLLM Proxy
        run: |
          ./scripts/litellm_run.sh
        env:
          AZURE_API_BASE: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_BASE }}
          AZURE_API_KEY: ${{ secrets.AZURE_GROUP_SUBSCRIPTION_API_KEY }}

      - name: Prepare RAG dataset
        run: |
          set -euo pipefail
          cd examples/rag
          mkdir -p data
          uv run gdown --fuzzy "https://drive.google.com/file/d/1Pq4Ag8zVoN8gUtLu0LcBfY35Dm5zL0hq/view?usp=drive_link" -O data/dataset_tiny.parquet
          uv run gdown --fuzzy "https://drive.google.com/file/d/1REXCpRLbeZu1KfWWKhIGEQe_WNHUOBkS/view?usp=drive_link" -O data/chunks_candidate_tiny.pkl
          uv run gdown --fuzzy "https://drive.google.com/file/d/1f6P-h_8KSRhe5pqDHWbRQWvUhTygfZ-c/view?usp=drive_link" -O data/index_hnsw_faiss_n32e40_tiny.index

      - name: Run WIKI Retriever MCP Server
        run: |
          set -euo pipefail
          cd examples/rag
          uv run python wiki_retriever_mcp.py &
          for i in {1..20}; do
            sleep 5
            if nc -z localhost 8099; then
              echo "MCP server is up!"
              exit 0
            else
              echo "Waiting for MCP server to start..."
            fi
          done
          echo "MCP server failed to start within expected time."
          exit 1

      - name: Run vLLM Server
        run: |
          set -euo pipefail
          source .venv/bin/activate
          vllm serve Qwen/Qwen2.5-1.5B-Instruct \
            --enable-auto-tool-choice \
            --tool-call-parser hermes \
            --port 8000 &

          VLLM_READY=0
          for i in {1..100}; do
            if curl -sSf http://localhost:8000/v1/models > /dev/null 2>&1; then
              echo "vLLM server is ready!"
              VLLM_READY=1
              break
            fi
            echo "Waiting for vLLM server to be ready... (${i})"
            sleep 5
          done
          if [[ "$VLLM_READY" != "1" ]]; then
            echo "vLLM server failed to start!"
            exit 1
          fi

      - name: Run RAG Sanity check
        run: |
          set -ex
          source .venv/bin/activate
          cd examples/rag
          uv run python rag_agent.py
        shell: bash

      - name: Stop vLLM Server
        run: |
          set -euo pipefail
          pkill -f vllm
          for i in {1..60}; do
            if ! pgrep -f vllm; then
              break
            fi
            sleep 5
          done

      - name: RAG training
        run: |
          set -ex
          source .venv/bin/activate
          cd examples/rag
          ../../scripts/restart_ray.sh
          sleep 5
          PYTHONUNBUFFERED=1 python train_rag.py fast
          sleep 10
        shell: bash
        env:
          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}
        id: rag_train

      - name: Validate RAG training
        run: |
          set -ex
          # Allow up to 5 rollouts to fail to produce rewards
          uv run scripts/validate_example_wandb.py ${{ steps.rag_train.outputs.project_name }} ${{ steps.rag_train.outputs.run_name }} --reward-tolerance 5
        env:
          WANDB_BASE_URL: ${{ secrets.MSR_WANDB_BASE_URL }}
          WANDB_API_KEY: ${{ secrets.MSR_WANDB_API_KEY }}