# Agora Cosmica self-host.
#
# Quickstart:
#   1. cp .env.example .env       (edit values you want to override)
#   2. docker compose up -d
#   3. open http://localhost:8080
#
# By default the app pulls a prebuilt image from GHCR. To build from source
# instead, comment the `image:` line and uncomment the `build:` block below.
#
# Local Mode (v1.1.0+) adds three sibling services:
#   - tts-kokoro : English TTS (Kokoro, 82M params, CPU or GPU)
#   - tts-qwen   : German TTS (Qwen3-TTS, CUDA-only) — under the `nvidia` profile
#   - stt-whisper: Speech-to-text (Whisper, faster-whisper backend)
#
# Profiles:
#   docker compose up                       → app + Kokoro + Whisper
#   docker compose --profile nvidia up      → adds tts-qwen for full DE local
#   (Apple Silicon: run scripts/setup-local-tts-apple.sh for native MLX Qwen)

services:
  app:
    image: ghcr.io/chipmates/agoracosmica:latest
    # build:
    #   context: ./client
    #   dockerfile: Dockerfile
    container_name: agoracosmica
    restart: unless-stopped
    ports:
      - "${AGORA_HOST_PORT:-8080}:8080"
    environment:
      AGORA_MEDIA_BASE_URL: ${AGORA_MEDIA_BASE_URL:-https://media.agoracosmica.org}
      AGORA_AUDIO_API_URL: ${AGORA_AUDIO_API_URL:-}
      AGORA_LOCAL_TTS_KOKORO_URL: ${AGORA_LOCAL_TTS_KOKORO_URL:-http://localhost:8880}
      AGORA_LOCAL_TTS_QWEN_URL: ${AGORA_LOCAL_TTS_QWEN_URL:-http://localhost:8887}
      AGORA_LOCAL_STT_URL: ${AGORA_LOCAL_STT_URL:-http://localhost:8000}
    healthcheck:
      test: ["CMD", "wget", "-q", "-O", "-", "http://127.0.0.1:8080/healthz"]
      interval: 30s
      timeout: 3s
      retries: 3
      start_period: 5s
    # Optional same-origin media mirror. Drop a content tree into ./media/
    # and set AGORA_MEDIA_BASE_URL=/media to serve the catalog from your own
    # box instead of the public CDN.
    # volumes:
    #   - ./media:/usr/share/nginx/html/media:ro

  tts-kokoro:
    image: ghcr.io/chipmates/agoracosmica-tts-kokoro:latest
    # build:
    #   context: ./docker/tts-kokoro
    container_name: agoracosmica-tts-kokoro
    restart: unless-stopped
    ports:
      - "${AGORA_TTS_KOKORO_PORT:-8880}:8880"
    healthcheck:
      test: ["CMD", "wget", "-q", "-O", "-", "http://127.0.0.1:8880/v1/models"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 60s

  stt-whisper:
    image: ghcr.io/chipmates/agoracosmica-stt-whisper:latest
    # build:
    #   context: ./docker/stt-whisper
    container_name: agoracosmica-stt-whisper
    restart: unless-stopped
    ports:
      - "${AGORA_STT_PORT:-8000}:8000"
    environment:
      WHISPER__MODEL: ${AGORA_WHISPER_MODEL:-deepdml/faster-whisper-large-v3-turbo-ct2}
      WHISPER__INFERENCE_DEVICE: ${AGORA_WHISPER_DEVICE:-auto}
    volumes:
      # Persist the HuggingFace model cache across restarts so the pre-pull
      # in entrypoint.sh runs once. Without this, every `docker compose down`
      # would force a fresh ~1.5 GB download on the next `up`.
      - whisper-cache:/home/ubuntu/.cache/huggingface
    healthcheck:
      test: ["CMD", "wget", "-q", "-O", "-", "http://127.0.0.1:8000/v1/models"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 60s

  tts-qwen:
    image: ghcr.io/chipmates/agoracosmica-tts-qwen-cuda:latest
    # build:
    #   context: ./docker/tts-qwen-cuda
    container_name: agoracosmica-tts-qwen
    restart: unless-stopped
    ports:
      - "${AGORA_TTS_QWEN_PORT:-8887}:8887"
    profiles: ["nvidia"]
    environment:
      VOICES_R2_BASE: ${AGORA_VOICES_R2_BASE:-https://media.agoracosmica.org/voices}
      GPU_MEMORY_UTILIZATION: ${AGORA_QWEN_GPU_MEM:-0.4}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    healthcheck:
      test: ["CMD", "wget", "-q", "-O", "-", "http://127.0.0.1:8887/health"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 120s

volumes:
  whisper-cache: