docker-compose.yml

services:
  api:
    build:
      context: .
      dockerfile: Dockerfile
      args:
        # Baked into frontend at `npm run build` time (Vite inlines env).
        # api still reads PRECIS_API_KEY at runtime via env_file / environment.
        VITE_API_BASE_URL: ""  # same-origin in Docker (frontend served by FastAPI)
        PRECIS_API_KEY: ${PRECIS_API_KEY:-}
        VITE_API_KEY: ${VITE_API_KEY:-${PRECIS_API_KEY:-}}
    container_name: precis-api
    ports:
      - "8000:8000"
    env_file:
      - .env
    environment:
      # Host Ollama (Option A) — blazing fast GPU, no 2.5GB re-download.
      # api inside Docker talks to host via host.docker.internal:11434.
      # Host must have `ollama serve` running (see README). On Windows Docker Desktop this resolves automatically.
      # If host Ollama was started with default 127.0.0.1 binding, set OLLAMA_HOST=0.0.0.0 on host before `ollama serve`.
      PORT: ${PORT:-8000}
      OLLAMA_BASE_URL: http://host.docker.internal:11434
      DEFAULT_MODEL: ${DEFAULT_MODEL:-phi4-mini:latest}
      AVAILABLE_MODELS: ${AVAILABLE_MODELS:-phi4-mini:latest}
      PRECIS_ALLOWED_ORIGINS: ${PRECIS_ALLOWED_ORIGINS:-http://localhost:5173,http://localhost:8000,http://localhost:7860,https://*.hf.space,https://*.huggingface.co}
      PRECIS_API_KEY: ${PRECIS_API_KEY:?PRECIS_API_KEY must be set in .env}
      OLLAMA_KEEP_ALIVE: ${OLLAMA_KEEP_ALIVE:-30m}
    extra_hosts:
      - "host.docker.internal:host-gateway"
    volumes:
      # Mount .env read-only so config.py hot-reload isn't needed — container restarts pick up changes
      - type: bind
        source: ./.env
        target: /app/.env
        read_only: true
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-fsS", "http://localhost:${PORT:-8000}/health"]
      interval: 15s
      timeout: 5s
      retries: 5
      start_period: 20s

volumes: {}