changemaker.lite/media-manager/docker-compose.yml

services:
  # PostgreSQL Database
  postgres:
    image: postgres:15-alpine
    restart: unless-stopped
    environment:
      POSTGRES_DB: library
      POSTGRES_USER: mediamanager
      # REQUIRED: Set POSTGRES_PASSWORD in .env file
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?POSTGRES_PASSWORD must be set in .env}
    volumes:
      - postgres-data:/var/lib/postgresql/data
    ports:
      # Bind to localhost only to prevent external access
      - "127.0.0.1:5432:5432"
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U mediamanager -d library"]
      interval: 5s
      timeout: 5s
      retries: 5
    networks:
      - media-network

  # API Server
  api:
    build:
      context: .
      dockerfile: packages/api/Dockerfile
    ports:
      - "3001:3001"
    volumes:
      - ./data:/app/data
      # Docker socket for container management (JoyCaption start/stop)
      - /var/run/docker.sock:/var/run/docker.sock
      # Mount local library for organization (read-write for scripts)
      - /media/bunker-admin/Internal/plex/xxx/media/local:/media/local
      # Mount studios for scanning (includes OnlyFans as subdirectory)
      - /media/bunker-admin/Internal/plex/xxx/media/local/studios:/media/studios
      # Mount public curated content (read-write for copy operations)
      - /media/bunker-admin/Internal/plex/xxx/media/public/curated:/media/public/curated
      # Mount playback directory for compilation output (read-write)
      - /media/bunker-admin/Internal/plex/xxx/media/public/playback:/media/playback
      # Mount compilations directory for symlinks (read-write)
      - /media/bunker-admin/Internal/plex/xxx/media/public/compilations:/media/compilations
      # Mount videos directory for public videos (read-write for copy operations)
      - /media/bunker-admin/Internal/plex/xxx/media/public/videos:/media/videos
      # Mount quickies directory for short-form videos (read-write)
      - /media/bunker-admin/Internal/plex/xxx/media/public/quickies:/media/quickies
      # Mount gifs directory for compilation scripts
      - /media/bunker-admin/Internal/plex/xxx/media/local/gifs:/media/gifs
      # Mount inbox directory for downloaded files
      - /media/bunker-admin/Internal/plex/xxx/media/local/inbox:/media/inbox
      # Mount scripts directory (read-only)
      - ./scripts:/app/scripts:ro
      # Mount anime models directory for model status checking (read-only)
      - ./models/anime-pipe:/app/models/anime-pipe:ro
      # Mount JoyTag models directory for tag seeding (read-only)
      - ./models/joytag:/app/models/joytag:ro
    depends_on:
      postgres:
        condition: service_healthy
    extra_hosts:
      - "host.docker.internal:host-gateway"
    environment:
      - DATABASE_URL=postgresql://mediamanager:${POSTGRES_PASSWORD}@postgres:5432/library
      - DATA_DIR=/app/data
      - MEDIA_PATH=/media
      - PUBLIC_DIR=/media/public/curated
      - SOURCE_LIBRARY=/media/studios
      - GIFS_PATH=/media/gifs
      - INBOX_PATH=/media/inbox
      - PLAYBACK_PATH=/media/playback
      - COMPILATIONS_PATH=/media/compilations
      - VIDEOS_PATH=/media/videos
      - QUICKIES_PATH=/media/quickies
      - SCRIPTS_PATH=/app/scripts
      - PORT=3001
      # JWT Authentication - REQUIRED: Set JWT_SECRET in .env file
      - JWT_SECRET=${JWT_SECRET:?JWT_SECRET must be set in .env}
      # Initial admin user (created on first startup if no admin exists)
      - INITIAL_ADMIN_EMAIL=${INITIAL_ADMIN_EMAIL:-}
      - INITIAL_ADMIN_PASSWORD=${INITIAL_ADMIN_PASSWORD:-}
      # SMTP Configuration (for email verification and password reset)
      - SMTP_HOST=${SMTP_HOST:-}
      - SMTP_PORT=${SMTP_PORT:-587}
      - SMTP_SECURE=${SMTP_SECURE:-false}
      - SMTP_USER=${SMTP_USER:-}
      - SMTP_PASS=${SMTP_PASS:-}
      - SMTP_FROM=${SMTP_FROM:-noreply@example.com}
      - SMTP_FROM_NAME=${SMTP_FROM_NAME:-Media Manager}
      - EMAIL_VERIFICATION_EXPIRY_HOURS=${EMAIL_VERIFICATION_EXPIRY_HOURS:-24}
      - PASSWORD_RESET_EXPIRY_HOURS=${PASSWORD_RESET_EXPIRY_HOURS:-1}
      - APP_BASE_URL=${APP_BASE_URL:-http://localhost:3080}
      # CORS allowed origins (comma-separated list of allowed domains)
      - ALLOWED_ORIGINS=${ALLOWED_ORIGINS:-http://localhost:3080,http://localhost:8080}
      # Timezone
      - TZ=${TZ:-UTC}
      # MaxMind GeoIP Configuration (optional - for session geographic data)
      # Register free at: https://www.maxmind.com/en/geolite2/signup
      - MAXMIND_ACCOUNT_ID=${MAXMIND_ACCOUNT_ID:-}
      - MAXMIND_LICENSE_KEY=${MAXMIND_LICENSE_KEY:-}
      # Content Safety (Ollama) Configuration
      - OLLAMA_BASE_URL=${OLLAMA_BASE_URL:-http://ollama:11434}
      - OLLAMA_MODEL=${OLLAMA_MODEL:-llama-guard3:1b}
      - SAFETY_CHECK_ENABLED=${SAFETY_CHECK_ENABLED:-true}
      # Digest Feature Configuration
      - DIGEST_OLLAMA_URL=${DIGEST_OLLAMA_URL:-http://ollama:11434}
      - DIGEST_WHISPER_URL=${DIGEST_WHISPER_URL:-http://whisper:5000}
      # Use abliterated qwen3 models for uncensored scene analysis
      - DIGEST_VISION_MODEL=${DIGEST_VISION_MODEL:-huihui_ai/qwen3-vl-abliterated:2b}
      - DIGEST_TEXT_MODEL=${DIGEST_TEXT_MODEL:-huihui_ai/qwen3-abliterated:4b}
      - DIGEST_FRAME_INTERVAL=${DIGEST_FRAME_INTERVAL:-30}
      - DIGEST_TEMP_DIR=/app/data/digest_temp
      # JoyCaption Vision Backend Configuration (alternative to Ollama vision)
      - DIGEST_VISION_BACKEND=${DIGEST_VISION_BACKEND:-ollama}
      - JOYCAPTION_URL=${JOYCAPTION_URL:-http://joycaption:8080}
      - JOYCAPTION_MODEL=${JOYCAPTION_MODEL:-llama-joycaption-beta-one-hf-llava}
      - JOYCAPTION_FALLBACK_TO_OLLAMA=${JOYCAPTION_FALLBACK_TO_OLLAMA:-true}
      # Face Recognition Service Configuration
      - FACE_RECOGNITION_URL=${FACE_RECOGNITION_URL:-http://face-recognition:5001}
      # YOLO11 Person Detection Service Configuration
      - YOLO_DETECTION_URL=${YOLO_DETECTION_URL:-http://yolo-detection:5002}
      # JoyTag Visual Tagging Service Configuration
      - JOYTAG_URL=${JOYTAG_URL:-http://joytag:5003}
      - JOYTAG_ENABLED=${JOYTAG_ENABLED:-true}
      - JOYTAG_THRESHOLD=${JOYTAG_THRESHOLD:-0.35}
      - JOYTAG_TOP_K=${JOYTAG_TOP_K:-50}
      # Scene Detection Service Configuration
      - SCENE_DETECTION_URL=${SCENE_DETECTION_URL:-http://scene-detection:5004}
      - SCENE_DETECTION_ENABLED=${SCENE_DETECTION_ENABLED:-true}
      # TransNetV2 Neural Scene Detection Configuration (fast pipeline)
      - TRANSNET_URL=${TRANSNET_URL:-http://transnetv2:5005}
      - TRANSNET_ENABLED=${TRANSNET_ENABLED:-true}
      - TRANSNET_THRESHOLD=${TRANSNET_THRESHOLD:-0.5}
      # CLIP Embeddings Configuration (fast pipeline diversity scoring)
      - CLIP_EMBEDDINGS_URL=${CLIP_EMBEDDINGS_URL:-http://clip-embeddings:5006}
      - CLIP_EMBEDDINGS_ENABLED=${CLIP_EMBEDDINGS_ENABLED:-true}
      - CLIP_FRAME_BATCH_SIZE=${CLIP_FRAME_BATCH_SIZE:-512}
      # Container Lifecycle Configuration
      # Set to 'true' when AI containers run continuously (models fit in VRAM)
      - DISABLE_CONTAINER_LIFECYCLE=${DISABLE_CONTAINER_LIFECYCLE:-false}
      # Combined Scene Detection Configuration
      - ENABLE_COMBINED_SCENE_DETECTION=${ENABLE_COMBINED_SCENE_DETECTION:-true}
      - SCENE_MERGE_STRATEGY=${SCENE_MERGE_STRATEGY:-weighted}
      - SCENE_MERGE_TOLERANCE=${SCENE_MERGE_TOLERANCE:-0.75}
      - SCENE_DETECTION_PARALLEL=${SCENE_DETECTION_PARALLEL:-false}
      - SCENE_WEIGHT_TRANSNET=${SCENE_WEIGHT_TRANSNET:-1.0}
      - SCENE_WEIGHT_PYSCENEDETECT=${SCENE_WEIGHT_PYSCENEDETECT:-0.8}
      - SCENE_WEIGHT_CLIP=${SCENE_WEIGHT_CLIP:-0.9}
      - CLIP_BOUNDARY_FRAME_INTERVAL=${CLIP_BOUNDARY_FRAME_INTERVAL:-0.5}
      - CLIP_BOUNDARY_SIMILARITY_THRESHOLD=${CLIP_BOUNDARY_SIMILARITY_THRESHOLD:-0.7}
      - CLIP_BOUNDARY_MIN_GAP=${CLIP_BOUNDARY_MIN_GAP:-2.0}
      # Payment System IMAP Configuration
      - PAYMENT_IMAP_HOST=${PAYMENT_IMAP_HOST:-}
      - PAYMENT_IMAP_PORT=${PAYMENT_IMAP_PORT:-993}
      - PAYMENT_IMAP_USER=${PAYMENT_IMAP_USER:-}
      - PAYMENT_IMAP_PASS=${PAYMENT_IMAP_PASS:-}
      - PAYMENT_IMAP_TLS=${PAYMENT_IMAP_TLS:-true}
      - PAYMENT_POLL_INTERVAL_MS=${PAYMENT_POLL_INTERVAL_MS:-60000}
      - PAYMENT_PROCESSED_FOLDER=${PAYMENT_PROCESSED_FOLDER:-Processed}
      # PornDB API Configuration (metadata enrichment)
      - PORNDB_API_KEY=${PORNDB_API_KEY:-}
      # Inbox File Watcher Configuration (real-time auto-digest)
      - FILE_WATCHER_ENABLED=${FILE_WATCHER_ENABLED:-false}
      - FILE_WATCHER_DEBOUNCE_MS=${FILE_WATCHER_DEBOUNCE_MS:-2000}
      # FFmpeg parallel encoding (default 1 sequential — NVENC is fastest single-stream)
      - FFMPEG_ENCODE_CONCURRENCY=${FFMPEG_ENCODE_CONCURRENCY:-1}
    # NVIDIA GPU support for NVENC encoding
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu, video]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:3001/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 30s

  # Admin UI - directly exposed for local dev
  admin:
    build:
      context: .
      dockerfile: packages/admin/Dockerfile
    ports:
      - "8080:3000"
    environment:
      - VITE_API_URL=http://api:3001
    depends_on:
      - api
    networks:
      - media-network
    restart: unless-stopped

  # Public gallery
  public:
    build:
      context: .
      dockerfile: packages/public/Dockerfile
      args:
        - VITE_API_URL=${VITE_API_URL:-http://api:3001}
        - VITE_VIDEO_URL=${VITE_VIDEO_URL:-http://localhost:8081}
    ports:
      - "3080:80"
    environment:
      - VITE_API_URL=${VITE_API_URL:-http://api:3001}
      - VITE_VIDEO_URL=${VITE_VIDEO_URL:-http://localhost:8081}
    depends_on:
      - api
      - nginx
    networks:
      - media-network
    restart: unless-stopped

  # Nginx for video streaming (public and library content)
  nginx:
    image: nginx:alpine
    ports:
      - "8081:80"
    environment:
      - APP_BASE_URL=${APP_BASE_URL:-http://localhost:3080}
    volumes:
      - /media/bunker-admin/Internal/plex/xxx/media/public/curated:/media/public/curated:ro
      - /media/bunker-admin/Internal/plex/xxx/media/public/playback:/media/playback:ro
      - /media/bunker-admin/Internal/plex/xxx/media/public/compilations:/media/compilations:ro
      - /media/bunker-admin/Internal/plex/xxx/media/public/videos:/media/videos:ro
      - /media/bunker-admin/Internal/plex/xxx/media/public/quickies:/media/quickies:ro
      - /media/bunker-admin/Internal/plex/xxx/media/local:/media/local:ro
      - ./data/thumbnails:/media/thumbnails:ro
      - ./data/ads:/media/ads:ro
      - ./data/gallery:/media/gallery:ro
      - ./data/digest_frames:/media/digest_frames:ro
      - ./nginx/nginx.conf:/etc/nginx/nginx.conf.template:ro
    command: /bin/sh -c "envsubst '$$APP_BASE_URL' < /etc/nginx/nginx.conf.template > /etc/nginx/nginx.conf && nginx -g 'daemon off;'"
    networks:
      - media-network
    restart: unless-stopped

  # Ollama LLM Server
  # For safety: Uses Llama Guard 3 1B (~1.6GB) for content moderation
  # For digest: Uses qwen3-vl (vision) and qwen3 (text) for video analysis
  # Models are automatically pulled on first startup based on OLLAMA_MODELS env var
  ollama:
    profiles: ["ai"]
    build:
      context: ./ollama
      dockerfile: Dockerfile
    ports:
      - "11435:11434"
    volumes:
      - ollama-data:/root/.ollama
    environment:
      - OLLAMA_KEEP_ALIVE=24h
      # Comma-separated list of models to pull on startup
      # Abliterated qwen3 models for uncensored analysis
      - OLLAMA_MODELS=${OLLAMA_MODELS:-llama-guard3:1b,huihui_ai/qwen3-vl-abliterated:2b,huihui_ai/qwen3-abliterated:4b}
      - OLLAMA_CONTEXT_LENGTH=${OLLAMA_CONTEXT_LENGTH:-4096}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:11434/"]
      interval: 30s
      timeout: 10s
      retries: 5
      start_period: 180s

  # Whisper Transcription Service
  # Uses faster-whisper for GPU-accelerated speech recognition
  whisper:
    profiles: ["ai"]
    build:
      context: ./whisper
      dockerfile: Dockerfile
    ports:
      - "5000:5000"
    volumes:
      - whisper-models:/root/.cache
    environment:
      - WHISPER_MODEL=${WHISPER_MODEL:-base}
      - WHISPER_DEVICE=${WHISPER_DEVICE:-cuda}
      - WHISPER_COMPUTE=${WHISPER_COMPUTE:-float16}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:5000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s

  # JoyCaption Vision Model Server
  # Uses llama.cpp server with JoyCaption GGUF for image captioning
  # Requires model files in ./models directory:
  #   - llama-joycaption-beta-one-hf-llava.Q2_K.gguf (~2GB)
  #   - llama-joycaption-beta-one-llava-mmproj-model-f16.gguf (~600MB)
  # Download from: https://huggingface.co/Mungert/llama-joycaption-beta-one-hf-llava-GGUF
  joycaption:
    profiles: ["ai"]
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    ports:
      - "11436:8080"
    volumes:
      - ./models:/models:ro
    command:
      - "--model"
      - "${JOYCAPTION_MODEL_PATH:-/models/llama-joycaption-beta-one-hf-llava.Q2_K.gguf}"
      - "--mmproj"
      - "${JOYCAPTION_MMPROJ_PATH:-/models/llama-joycaption-beta-one-llava-mmproj-model-f16.gguf}"
      - "--host"
      - "0.0.0.0"
      - "--port"
      - "8080"
      - "--ctx-size"
      - "${JOYCAPTION_CTX_SIZE:-4096}"
      - "--n-gpu-layers"
      - "${JOYCAPTION_GPU_LAYERS:-999}"
      - "--parallel"
      - "${JOYCAPTION_PARALLEL:-1}"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8080/health"]
      interval: 30s
      timeout: 10s
      retries: 5
      start_period: 120s

  # Face Recognition Service
  # Uses InsightFace with ONNX Runtime for face detection and embedding generation
  # Provides face counting, gender detection, and performer recognition
  face-recognition:
    profiles: ["ai"]
    build:
      context: ./face-recognition
      dockerfile: Dockerfile
    ports:
      - "5001:5001"
    volumes:
      - face-models:/root/.insightface
    environment:
      - CUDA_VISIBLE_DEVICES=0
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:5001/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s

  # YOLO11 Person Detection Service (for performer counting)
  yolo-detection:
    profiles: ["ai"]
    build:
      context: ./yolo-detection
      dockerfile: Dockerfile
    ports:
      - "5002:5002"
    environment:
      - MODEL=yolo11s.pt
      - CONF_THRESHOLD=0.5
      - DEVICE=0
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:5002/health')"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 120s

  # JoyTag Visual Tagging Service (for multi-label image tagging)
  # Uses JoyTag ViT-B/16 for 5000+ Danbooru-style tags with confidence scores
  # Model auto-downloaded from HuggingFace on first startup (~366MB ONNX)
  joytag:
    profiles: ["ai"]
    build:
      context: ./joytag
      dockerfile: Dockerfile
    ports:
      - "5003:5003"
    environment:
      - CUDA_VISIBLE_DEVICES=0
      - CUDA_DEVICE=0
      - MODEL_DIR=/models/joytag
      - PORT=5003
    volumes:
      - ./models:/models
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:5003/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 120s

  # Scene Detection Service (for hard cut detection in videos)
  # Uses PySceneDetect to find scene boundaries for cleaner clip transitions
  # CPU-based - no GPU required
  scene-detection:
    profiles: ["ai"]
    build:
      context: ./scene-detection
      dockerfile: Dockerfile
    ports:
      - "5004:5004"
    volumes:
      # Mount media directories (read-only) for video access
      - /media/bunker-admin/Internal/plex/xxx/media/local:/media/local:ro
      - /media/bunker-admin/Internal/plex/xxx/media/public:/media/public:ro
      # Mount inbox for videos pending processing
      - /media/bunker-admin/Internal/plex/xxx/media/local/inbox:/media/inbox:ro
      # Mount digest temp for chunked video processing
      - ./data/digest_temp:/app/data/digest_temp:ro
    environment:
      - SCENE_CONTENT_THRESHOLD=${SCENE_CONTENT_THRESHOLD:-27.0}
      - SCENE_MIN_LENGTH=${SCENE_MIN_LENGTH:-15}
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:5004/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 10s

  # TransNetV2 Neural Shot Boundary Detection Service
  # Uses TransNetV2 for faster and more accurate scene detection
  # Especially good at detecting soft transitions (dissolves, fades)
  # Used by fast digest pipeline for scene-based keyframe extraction
  transnetv2:
    profiles: ["ai"]
    build:
      context: ./transnetv2
      dockerfile: Dockerfile
    ports:
      - "5005:5005"
    volumes:
      # Mount media directories (read-only) for video access
      - /media/bunker-admin/Internal/plex/xxx/media/local:/media/local:ro
      - /media/bunker-admin/Internal/plex/xxx/media/public:/media/public:ro
      - /media/bunker-admin/Internal/plex/xxx/media/local/inbox:/media/inbox:ro
      # Mount digest temp for chunked video processing
      - ./data/digest_temp:/app/data/digest_temp:ro
    environment:
      - TRANSNET_THRESHOLD=${TRANSNET_THRESHOLD:-0.5}
      - TRANSNET_MIN_SCENE_LENGTH=${TRANSNET_MIN_SCENE_LENGTH:-1.0}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:5005/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s

  # CLIP Embeddings Service for Visual Diversity Scoring
  # Uses OpenAI CLIP ViT-B/32 to extract 512-D visual embeddings
  # Enables diversity-based clip selection without LLM calls
  # Used by fast digest pipeline for visual similarity measurement
  clip-embeddings:
    profiles: ["ai"]
    build:
      context: ./clip-embeddings
      dockerfile: Dockerfile
    ports:
      - "5006:5006"
    environment:
      - CLIP_MODEL=${CLIP_MODEL:-ViT-B/32}
      # GPU batch size for embedding extraction (increase if VRAM allows, decrease if OOM)
      - CLIP_GPU_BATCH_SIZE=${CLIP_GPU_BATCH_SIZE:-512}
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    networks:
      - media-network
    restart: unless-stopped
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:5006/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s

  # Newt client for Pangolin tunnel
  newt:
    image: fosrl/newt
    container_name: newt
    restart: unless-stopped
    environment:
      - PANGOLIN_ENDPOINT=${PANGOLIN_ENDPOINT}
      - NEWT_ID=${NEWT_ID}
      - NEWT_SECRET=${NEWT_SECRET}
    networks:
      - media-network

volumes:
  postgres-data:
  ollama-data:
  whisper-models:
  face-models:

networks:
  media-network:
    external: true