services:
  wildecho-api:
    build:
      context: .
      dockerfile: Dockerfile
    image: wildecho-api:local
    container_name: wildecho-api
    restart: unless-stopped
    ports:
      - "${WILDECHO_PORT:-8000}:8000"

    volumes:
      # Model weights live on the host and are mounted read-only. Populate this
      # directory once with:  python scripts/download_model.py
      # Weights are ~390 MiB, Apache-2.0, copyright Google LLC, and are never
      # committed to git or baked into the image.
      - ./models:/app/models:ro

    environment:
      WILDECHO_MODEL_PATH: /app/models/perch_v2.onnx
      WILDECHO_TAXONOMY_PATH: /app/data/taxonomy.csv

      # Per-IP rate limit. The default is a conservative free-tier figure. Raise it
      # for private use; see the README for the reverse-proxy caveat.
      WILDECHO_RATE_LIMIT: "${WILDECHO_RATE_LIMIT:-20/hour}"
      WILDECHO_RATE_LIMIT_ENABLED: "${WILDECHO_RATE_LIMIT_ENABLED:-true}"

      # Permissive by default so a mobile app works without a proxy. Set this to
      # your app's origins before exposing the service publicly.
      WILDECHO_CORS_ORIGINS: "${WILDECHO_CORS_ORIGINS:-*}"

      WILDECHO_MAX_UPLOAD_BYTES: "${WILDECHO_MAX_UPLOAD_BYTES:-26214400}"
      WILDECHO_MAX_DURATION_SECONDS: "${WILDECHO_MAX_DURATION_SECONDS:-300}"
      WILDECHO_LOW_CONFIDENCE_THRESHOLD: "${WILDECHO_LOW_CONFIDENCE_THRESHOLD:-0.3}"
      WILDECHO_TOP_K: "${WILDECHO_TOP_K:-10}"
      WILDECHO_LOG_LEVEL: "${WILDECHO_LOG_LEVEL:-INFO}"

    healthcheck:
      test: ["CMD", "curl", "-fsS", "http://localhost:8000/v1/health"]
      interval: 30s
      timeout: 5s
      start_period: 40s
      retries: 3

    # Inference is CPU-bound. Uncomment and tune for your hardware; without a
    # limit one long clip can saturate every core on the host.
    # deploy:
    #   resources:
    #     limits:
    #       cpus: "2.0"
    #       memory: 2G