docker-compose.yml

name: amarin

services:
  searxng:
    image: searxng/searxng:latest
    ports:
      - "127.0.0.1:8888:8080"
    volumes:
      - ./searxng-settings:/etc/searxng
    environment:
      - SEARXNG_BASE_URL=http://searxng:8080/

  backend:
    build: ./backend
    extra_hosts:
      - "host.docker.internal:host-gateway"
    ports:
      - "127.0.0.1:8100:8000"
    volumes:
      - ./backend/data:/app/data
      - ./voices:/app/voices
      - ./:/app/source:ro
      # Docker socket removed (Phase 47 security hardening) — TTS provider
      # switching now requires restart via start.sh instead of runtime hot-swap.
      # The _switch_tts_containers function degrades gracefully without it.
    env_file:
      - .env
    environment:
      - DATABASE_URL=sqlite:///./data/amarin.db
      - SEARXNG_URL=http://searxng:8080/search
      - DIRECTUS_URL=${DIRECTUS_URL:-http://localhost:8055}
      - DIRECTUS_TOKEN=${DIRECTUS_TOKEN:-}
    depends_on:
      - searxng

  frontend:
    build: ./frontend
    ports:
      - "${FRONTEND_BIND:-0.0.0.0}:80:80"
    depends_on:
      - backend

  # ---------------------------------------------------------------------------
  # Speech-to-Text service (faster-whisper, GPU)
  # ---------------------------------------------------------------------------

  speech-service:
    build: ./speech-service
    container_name: amarin-speech
    ports:
      - "127.0.0.1:8900:8000"
    volumes:
      - whisper-cache:/root/.cache
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
      - PYTHONUNBUFFERED=1
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped

  # ---------------------------------------------------------------------------
  # Embedding service for semantic memory search
  # ---------------------------------------------------------------------------

  embedding:
    build: ./embedding-service
    container_name: amarin-embedding
    ports:
      - "127.0.0.1:8200:8200"
    volumes:
      - embedding-cache:/root/.cache
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped

  # ---------------------------------------------------------------------------
  # Perception Service — Qwen2.5-VL-3B, Auri's eyes (GPU)
  # Local vision perception: image description, scene/emotion analysis
  # ---------------------------------------------------------------------------
  perception:
    build: ./perception-service
    container_name: amarin-perception
    ports:
      - "127.0.0.1:8950:8000"
    volumes:
      - perception-cache:/root/.cache
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
      - PYTHONUNBUFFERED=1
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    profiles:
      - perception

  # ---------------------------------------------------------------------------
  # Kokoro TTS — Fast 82M param model with 67 voices + blending (GPU)
  # ---------------------------------------------------------------------------
  kokoro-tts:
    image: ghcr.io/remsky/kokoro-fastapi-gpu:latest
    container_name: amarin-kokoro-tts
    ports:
      - "127.0.0.1:8880:8880"
    environment:
      - USE_GPU=true
      - PYTHONUNBUFFERED=1
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    profiles:
      - kokoro  # Only start when explicitly requested

  # ---------------------------------------------------------------------------
  # Pocket TTS — Fast CPU-based TTS with voice cloning (100M model)
  # ---------------------------------------------------------------------------
  pocket-tts:
    build:
      context: ./pocket-tts-server
      dockerfile: Dockerfile
    container_name: amarin-pocket-tts
    ports:
      - "127.0.0.1:7870:7870"
    volumes:
      - ./voices:/app/voices
      - pocket-tts-cache:/root/.cache
    environment:
      - HF_TOKEN=${HF_TOKEN}
    env_file:
      - .env
    restart: unless-stopped
    profiles:
      - pocket-tts  # Only start when explicitly requested

  # ---------------------------------------------------------------------------
  # Inner Life LLM — CPU-only Qwen3-4B for autonomous background reflection
  # (upgraded from SmolLM3-3B — better reasoning, structured output, 32K context)
  # ---------------------------------------------------------------------------
  inner-llm:
    image: ghcr.io/ggml-org/llama.cpp:server
    container_name: amarin-inner-llm
    ports:
      - "127.0.0.1:8301:8080"
    volumes:
      - ./models:/models
    command: >
      -m /models/Qwen3-4B-Q4_K_M.gguf
      --host 0.0.0.0
      --port 8080
      --threads 6
      --ctx-size 8192
      --batch-size 512
    restart: unless-stopped
    profiles:
      - inner-life

  # ---------------------------------------------------------------------------
  # Local LLM — llama-cpp-python with CUDA (Dolphin 2.9.4 Llama 3.1 8B)
  # ---------------------------------------------------------------------------
  local-llm:
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    container_name: amarin-local-llm
    ports:
      - "127.0.0.1:8300:8080"
    volumes:
      - ./models:/models
    command: >
      -m /models/Dolphin3.0-Llama3.1-8B-Q4_K_M.gguf
      --host 0.0.0.0
      --port 8080
      --n-gpu-layers 25
      --ctx-size 8192
      --cache-type-k q8_0
      --cache-type-v q8_0
      --flash-attn on
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    profiles:
      - local-llm

  # ---------------------------------------------------------------------------
  # Orpheus TTS — Expressive 3B model with paralinguistic features (GPU)
  # Three containers: model downloader, llama.cpp inference, FastAPI audio server
  # ---------------------------------------------------------------------------
  orpheus-model-init:
    image: curlimages/curl:latest
    container_name: amarin-orpheus-model-init
    user: "1000:1000"
    volumes:
      - ./models:/models
    entrypoint: ["sh", "-c"]
    command:
      - |
        if [ ! -f /models/Orpheus-3b-FT-Q4_K_M.gguf ]; then
          echo "Downloading Orpheus 3B Q4_K_M (approx 2.4 GB)..."
          curl -L -o /models/Orpheus-3b-FT-Q4_K_M.gguf "https://huggingface.co/lex-au/Orpheus-3b-FT-Q4_K_M.gguf/resolve/main/Orpheus-3b-FT-Q4_K_M.gguf"
          echo "Download complete."
        else
          echo "Orpheus model already present."
        fi
    restart: "no"
    profiles:
      - orpheus

  orpheus-llm:
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    container_name: amarin-orpheus-llm
    volumes:
      - ./models:/models
    command: >
      -m /models/Orpheus-3b-FT-Q4_K_M.gguf
      --host 0.0.0.0
      --port 8080
      --n-gpu-layers 29
      --ctx-size 8192
      --flash-attn on
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    depends_on:
      orpheus-model-init:
        condition: service_completed_successfully
    profiles:
      - orpheus

  orpheus-tts:
    build:
      context: ./orpheus-tts
      dockerfile: Dockerfile.gpu
    container_name: amarin-orpheus-tts
    ports:
      - "127.0.0.1:5005:5005"
    environment:
      - ORPHEUS_API_URL=http://orpheus-llm:8080/v1/completions
      - ORPHEUS_MAX_TOKENS=8192
      - ORPHEUS_TEMPERATURE=0.4
      - ORPHEUS_TOP_P=0.9
      - ORPHEUS_SAMPLE_RATE=24000
      - ORPHEUS_MODEL_NAME=orpheus
      - ORPHEUS_PORT=5005
      - ORPHEUS_HOST=0.0.0.0
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    depends_on:
      - orpheus-llm
    profiles:
      - orpheus

  # ---------------------------------------------------------------------------
  # MOSS TTS — 1.7B param realtime synthesis with Auri's custom voice (GPU)
  # Voice reference: auri_voice_reference.wav (generated once by VoiceGenerator)
  # Generate voice: POST /voice/generate  |  Synthesize: POST /v1/audio/speech
  # ---------------------------------------------------------------------------
  moss-tts:
    build:
      context: ./moss-tts
      dockerfile: Dockerfile
    container_name: amarin-moss-tts
    ports:
      - "127.0.0.1:8885:8885"
    volumes:
      - ./moss-tts/data:/app/data
      - moss-hf-cache:/app/hf_cache
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
      - VOICE_REF_PATH=/app/data/auri_voice_reference.wav
      - REALTIME_MODEL=OpenMOSS-Team/MOSS-TTS-Realtime
      - VOICE_GEN_MODEL=OpenMOSS-Team/MOSS-VoiceGenerator
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    profiles:
      - moss  # Only start when explicitly requested

  # ---------------------------------------------------------------------------
  # Qwen3 TTS — 1.7B param expressive model with voice design + cloning (GPU)
  # Natural language instructions: whisper, emotion, pacing control
  # ---------------------------------------------------------------------------
  qwen3-tts:
    build:
      context: ./qwen3-tts
      dockerfile: Dockerfile
    container_name: amarin-qwen3-tts
    ports:
      - "127.0.0.1:8890:8880"
    volumes:
      - ./voices:/app/voices:ro
      - qwen3-model-cache:/root/.cache
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    restart: unless-stopped
    profiles:
      - qwen3

volumes:
  embedding-cache:
  perception-cache:
  pocket-tts-cache:
  whisper-cache:
  qwen3-model-cache:
  moss-hf-cache: