docker-compose.yml

services:
  # Frontend service (Vite)
  frontend:
    build:
      context: .
      target: dev
    ports:
      - "5173:5173"
    environment:
      - BUN_ENV=development
      - VITE_API_URL=http://localhost:3000
    command: ["bunx", "vite", "--host", "0.0.0.0"]
    develop:
      watch:
        - action: sync
          path: ./src/client
          target: /app/src/client
        - action: sync
          path: ./src/assets
          target: /app/src/assets
        - action: sync
          path: ./public
          target: /app/public
        - action: sync
          path: ./index.html
          target: /app/index.html
        - action: sync
          path: ./vite.config.ts
          target: /app/vite.config.ts
        - action: rebuild
          path: package.json
    networks:
      - app-network

  # --- AI Model Service (Ollama) --
  ollama:
    image: ollama/ollama:0.16.3 # This official Docker ollama image includes all necessary CUDA libraries for GPU support. Pulling latest version can cause changes in how certain models are run.
    #  You can try lightweight but unofficial CPU-only version alpine/ollama ((https://hub.docker.com/r/alpine/ollama) which is only around 70 MB. 
    container_name: ollama-shared # rename as needed
    ports:
      - "11434:11434"
    volumes:
      - ${OLLAMA_MODELS_PATH:-/home/cynth/ollama_models_shared}:/root/.ollama/models # Docker volume to bind mount pointing to a local shared models folder 
    # - ollama_models:/root/.ollama # persists downloaded models
    # NOTE: we should remove hard-coded models path and put env variable instead
    init: true # handles zombie processes, recommended for containers
    deploy: # deploy (lines 46-52) is more modern syntax vs. runtime: nvidia (legacy) but does the same thing
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all # if you have multiple GPUs
              capabilities: [gpu]
    environment:
      - OLLAMA_HOST=0.0.0.0 # listen on all interfaces
      - OLLAMA_PORT=11434
      - OLLAMA_NUM_PARALLEL=4 # number of parallel requests the model can handle
      - OLLAMA_MAX_LOADED_MODELS=2 # max models kept in memory simultaneously
      - OLLAMA_KEEP_ALIVE=-1 # keep models loaded forever (-1 = infinite)
      - OLLAMA_DEBUG=0
      - OLLAMA_CONTEXT_LENGTH=8192 # max context window size in tokens (affects VRAM)
      - NVIDIA_VISIBLE_DEVICES=all # which GPUs to expose to container 
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility # required GPU capabilities (ensures neccesary CUDA drivers loaded inside container)
    restart: unless-stopped
    healthcheck:
      test: [ "CMD", "ollama", "list" ] # checks if Ollama is responding
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 20s # gives Ollama time to start up
    networks:
      - app-network

  # --- Backend Service ---
  backend:
    build:
      context: .
      target: dev
    ports:
      - "3000:3000"
    volumes:
      - ./src/server/aiTest:/app/src/server/aiTest
    environment:
      - DATABASE_URL=postgresql://root:root@db:5432/test_db
      - NODE_ENV=development
      - PORT=3000
      - DB_HOST=db
      - DB_PORT=5432
      - DB_NAME=test_db
      - DB_USER=root
      - DB_PASSWORD=root
      - MODEL_URL=${MODEL_URL:-http://ollama:11434/v1/chat/completions} # Ollama OpenAI-compatible endpoint
      - TEXT2SQL_MODEL=${TEXT2SQL_MODEL:-arctic-text2sql:latest} # alternative model option
      # - TEXT2SQL_MODEL=${TEXT2SQL_MODEL:-distil-qwen3-4b:latest} # model for text-to-SQL tasks
      - AI_RESPONSE_MODEL=${AIRESP_MODEL:-qwen2.5-coder:7b} # main AI response model 
      - JUDGE_MODEL=${JUDGE_MODEL:-qwen2.5-coder:7b} # model for evaluating/validating responses
      # - JUDGE_MODEL=${JUDGE_MODEL:-qwen2.5-coder:14b} # larger judge model option (needs more VRAM)
      - MODEL_REGISTRY_URL=http://ollama:11434/api/tags
      - HARDWARE_TIER=4090-workstation

    command: ["bun", "run", "server"]
    develop:
      watch:
        - action: sync+restart
          path: ./src/server
          target: /app/src/server
          ignore:
            - "**/*.log"
            - "**/*.tmp"
        - action: sync+restart
          path: ./src/shared
          target: /app/src/shared
        - action: rebuild
          path: package.json
    depends_on:
      db:
        condition: service_started
      ollama:
        condition: service_healthy
    networks:
      - app-network

  # --- PostgreSQL Database ---
  db:
    image: postgres:16-alpine
    restart: always
    shm_size: 128mb
    environment:
      POSTGRES_USER: root
      POSTGRES_PASSWORD: root # NOTE: we'd remove these hard-coded passwords (since not really secure)
      POSTGRES_DB: test_db
    ports:
      - "5432:5432"
    volumes:
      - postgres_offline_data:/var/lib/postgresql/data
    networks:
      - app-network

  # --- pgAdmin GUI ---
  pgadmin:
    container_name: pgadmin4_container
    image: dpage/pgadmin4
    restart: always
    environment:
      PGADMIN_DEFAULT_EMAIL: admin@admin.com
      PGADMIN_DEFAULT_PASSWORD: root
    ports:
      - "5050:80"
    depends_on:
      - db
    volumes:
      - pgadmin_offline_data:/var/lib/pgadmin
    networks:
      - app-network

# --- Networks ---
networks:
  app-network:
    driver: bridge

# --- Volumes ---
volumes:
  postgres_offline_data:
  pgadmin_offline_data: