dgx-spark-playbooks/nvidia/txt2kg/assets/deploy/compose/docker-compose.yml

services:
  app:
    build:
      context: ../..
      dockerfile: deploy/app/Dockerfile
    ports:
      - '3001:3000'
    environment:
      # Neo4j configuration (ARM64 compatible - works with 64KB pages)
      - NEO4J_URI=bolt://neo4j:7687
      - NEO4J_USER=neo4j
      - NEO4J_PASSWORD=password123
      - GRAPH_DB_TYPE=neo4j
      # ArangoDB disabled - set to localhost to prevent DNS errors if accidentally accessed
      - ARANGODB_URL=http://localhost:8529
      - ARANGODB_DB=txt2kg
      - QDRANT_URL=http://qdrant:6333
      - VECTOR_DB_TYPE=qdrant
      - LANGCHAIN_TRACING_V2=true
      - SENTENCE_TRANSFORMER_URL=http://sentence-transformers:80
      - MODEL_NAME=all-MiniLM-L6-v2
      - EMBEDDINGS_API_URL=http://sentence-transformers:80
      - GRPC_SSL_CIPHER_SUITES=HIGH+ECDSA:HIGH+aRSA
      - NODE_TLS_REJECT_UNAUTHORIZED=0
      - OLLAMA_BASE_URL=http://ollama:11434/v1
      - OLLAMA_MODEL=llama3.1:8b
      - REMOTE_WEBGPU_SERVICE_URL=http://txt2kg-remote-webgpu:8083
      - NVIDIA_API_KEY=${NVIDIA_API_KEY:-}
      # Node.js timeout configurations for large model processing
      - NODE_OPTIONS=--max-http-header-size=80000
      - UV_THREADPOOL_SIZE=128
      - HTTP_TIMEOUT=1800000
      - REQUEST_TIMEOUT=1800000
    networks:
      - default
      - txt2kg-network
      - pinecone-net
    depends_on:
      neo4j:
        condition: service_healthy
      ollama:
        condition: service_started
      # Optional: sentence-transformers and entity-embeddings are only needed for vector search
      # Traditional graph search works without these services

  # Neo4j - ARM64 compatible graph database (works with 64KB page size kernel)
  neo4j:
    image: neo4j:5-community
    ports:
      - '7474:7474'  # HTTP
      - '7687:7687'  # Bolt
    environment:
      - NEO4J_AUTH=neo4j/password123
      - NEO4J_server_memory_heap_initial__size=512m
      - NEO4J_server_memory_heap_max__size=2G
    volumes:
      - neo4j_data:/data
      - neo4j_logs:/logs
    networks:
      - default
    restart: unless-stopped
    healthcheck:
      test: ["CMD-SHELL", "wget --no-verbose --tries=1 --spider http://localhost:7474 || exit 1"]
      interval: 15s
      timeout: 10s
      retries: 10
      start_period: 60s

  # ArangoDB disabled - doesn't support ARM64 with 64KB page size
  # Uncomment if using 4KB page kernel (linux-nvidia-6.14)
  # arangodb:
  #   image: arangodb:latest
  #   ports:
  #     - '8529:8529'
  #   environment:
  #     - ARANGO_NO_AUTH=1
  #   volumes:
  #     - arangodb_data:/var/lib/arangodb3
  #     - arangodb_apps_data:/var/lib/arangodb3-apps
  ollama:
    build:
      context: ../services/ollama
      dockerfile: Dockerfile
    image: ollama-custom:latest
    container_name: ollama-compose
    ports:
      - '11434:11434'
    volumes:
      - ollama_data:/root/.ollama
    environment:
      - NVIDIA_VISIBLE_DEVICES=all        # Make all GPUs visible to the container
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility  # Required capabilities for CUDA
      - CUDA_VISIBLE_DEVICES=0            # Use first GPU
      - OLLAMA_FLASH_ATTENTION=1          # Enable flash attention for better performance
      - OLLAMA_KEEP_ALIVE=30m             # Keep models loaded for 30 minutes
      - OLLAMA_NUM_PARALLEL=4             # Process 4 requests in parallel - DGX Spark has unified memory
      - OLLAMA_MAX_LOADED_MODELS=1        # Load only one model at a time to avoid VRAM contention
      - OLLAMA_KV_CACHE_TYPE=q8_0         # Reduce KV cache VRAM usage with minimal performance impact
      - OLLAMA_GPU_LAYERS=-1              # Force all layers on GPU
      - OLLAMA_LLM_LIBRARY=cuda           # Force CUDA backend
    networks:
      - default
    restart: unless-stopped
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    healthcheck:
      test: ["CMD", "ollama", "list"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s
  
  # Optional services for vector search (NOT required for traditional graph search)
  # Traditional graph search works with just: app, arangodb, and ollama
  sentence-transformers:
    build:
      context: ../services/sentence-transformers
      dockerfile: Dockerfile
    ports:
      - '8000:80'
    environment:
      - MODEL_NAME=all-MiniLM-L6-v2
    networks:
      - default
    restart: unless-stopped
    profiles:
      - vector-search  # Only start with: docker compose --profile vector-search up
  qdrant:
    image: qdrant/qdrant:latest
    container_name: qdrant
    ports:
      - "6333:6333"
      - "6334:6334"
    volumes:
      - qdrant_data:/qdrant/storage
    networks:
      - pinecone-net
    restart: unless-stopped
    profiles:
      - vector-search  # Only start with: docker compose --profile vector-search up
  qdrant-init:
    image: curlimages/curl:latest
    depends_on:
      - qdrant
    restart: "no"
    entrypoint: /bin/sh
    command:
      - -c
      - |
        echo 'Waiting for Qdrant to start...'
        sleep 5
        echo 'Checking if entity-embeddings collection exists...'
        RESPONSE=$(curl -s http://qdrant:6333/collections/entity-embeddings)
        if echo "$RESPONSE" | grep -q '"status":"ok"'; then
          echo 'entity-embeddings collection already exists'
        else
          echo 'Creating collection entity-embeddings...'
          curl -X PUT http://qdrant:6333/collections/entity-embeddings \
            -H 'Content-Type: application/json' \
            -d '{"vectors":{"size":384,"distance":"Cosine"}}'
          echo ''
          echo 'entity-embeddings collection created successfully'
        fi
        echo 'Checking if document-embeddings collection exists...'
        RESPONSE=$(curl -s http://qdrant:6333/collections/document-embeddings)
        if echo "$RESPONSE" | grep -q '"status":"ok"'; then
          echo 'document-embeddings collection already exists'
        else
          echo 'Creating collection document-embeddings...'
          curl -X PUT http://qdrant:6333/collections/document-embeddings \
            -H 'Content-Type: application/json' \
            -d '{"vectors":{"size":384,"distance":"Cosine"}}'
          echo ''
          echo 'document-embeddings collection created successfully'
        fi        
    networks:
      - pinecone-net
    profiles:
      - vector-search

volumes:
  neo4j_data:
  neo4j_logs:
  ollama_data:
  qdrant_data:
  # arangodb_data:
  # arangodb_apps_data:

networks:
  default:
    driver: bridge
  txt2kg-network:
    driver: bridge
  pinecone-net:
    name: pinecone