services:
  api:
    build:
      context: .
      args:
        # Swap to https://download.pytorch.org/whl/cu128 for the CUDA image.
        TORCH_INDEX: ${TORCH_INDEX:-https://download.pytorch.org/whl/cpu}
    environment:
      # The CLI backend needs an interactive Claude Code login, which a container does not have,
      # so the container runs on the API backend and wants a key.
      ANSWER_BACKEND: ${ANSWER_BACKEND:-api}
      DOCASSIST_API_KEY: ${DOCASSIST_API_KEY:-}
    volumes:
      - ./models:/app/models:ro          # hf download, see README
      - ./.cache:/app/.cache             # LLM summaries + HF cache, keyed by content hash
      - ./workspaces:/app/workspaces     # cloned and uploaded repositories
      - ./results:/app/results           # eval runs the dashboard reads
    ports:
      - "8000:8000"

  web:
    build: ./web
    depends_on:
      - api
    ports:
      - "5173:80"

# GPU: `docker compose --profile gpu up` after installing the NVIDIA container toolkit, and
# build with TORCH_INDEX pointing at the cu128 wheels. Without it the models run on CPU, which
# works but turns a ~500ms rerank into several seconds.
  api-gpu:
    profiles: ["gpu"]
    extends:
      service: api
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]