docker-compose.yml

services:
  # ==========================================================================
  # Cold tier: Neo4j (graph source of truth)
  # ==========================================================================
  neo4j:
    image: neo4j:5.26
    container_name: cogram-neo4j
    ports:
      - "7474:7474"
      - "7687:7687"
    environment:
      NEO4J_AUTH: neo4j/password
      NEO4J_PLUGINS: '["apoc"]'
      NEO4J_dbms_security_procedures_unrestricted: apoc.*
      NEO4J_server_memory_heap_max__size: ${NEO4J_HEAP:-512m}
      NEO4J_server_memory_pagecache_size: ${NEO4J_PAGECACHE:-256m}
    volumes:
      - neo4j_data:/data
      - neo4j_logs:/logs
    healthcheck:
      test: ["CMD-SHELL", "wget -O - --quiet http://localhost:7474 || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 12
    networks:
      - cogram

  # ==========================================================================
  # Hot tier: Redis (active subgraph cache)
  # ==========================================================================
  redis:
    image: redis:7-alpine
    container_name: cogram-redis
    command: ["redis-server", "/usr/local/etc/redis/redis.conf"]
    ports:
      - "6379:6379"
    volumes:
      - redis_data:/data
      - ./containers/redis/redis.conf:/usr/local/etc/redis/redis.conf:ro
    healthcheck:
      test: ["CMD", "redis-cli", "ping"]
      interval: 5s
      timeout: 3s
      retries: 5
    networks:
      - cogram

  # ==========================================================================
  # Warm tier: Postgres (Engram decision cache + audit + training_data)
  # ==========================================================================
  postgres:
    image: postgres:16-alpine
    container_name: cogram-postgres
    ports:
      - "5432:5432"
    environment:
      POSTGRES_USER: cogram
      POSTGRES_PASSWORD: cogram
      POSTGRES_DB: cogram
    volumes:
      - postgres_data:/var/lib/postgresql/data
      - ./containers/postgres/init.sql:/docker-entrypoint-initdb.d/init.sql:ro
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U cogram -d cogram"]
      interval: 5s
      timeout: 3s
      retries: 10
    networks:
      - cogram

  # ==========================================================================
  # Cogram MCP server (stdio + HTTP/SSE on :7800/mcp)
  # Image: pulled from ghcr.io by default. If unavailable (offline / fork),
  # falls back to building from local source via the build: block.
  # ==========================================================================
  cogram-mcp:
    image: ghcr.io/srk0102/cogram-mcp:latest
    pull_policy: missing
    build:
      context: .
      dockerfile: containers/cogram-mcp/Dockerfile
    container_name: cogram-mcp
    ports:
      - "7800:7800"
    environment:
      OPENAI_API_KEY: ${OPENAI_API_KEY:?set OPENAI_API_KEY in .env}
      OPENAI_BASE_URL: ${OPENAI_BASE_URL:-https://api.openai.com/v1}
      GRAPHITI_LLM_MODEL: ${GRAPHITI_LLM_MODEL:-gpt-4o-mini}
      ANNOTATOR_LLM_MODEL: ${ANNOTATOR_LLM_MODEL:-gpt-4o-mini}
      EMBEDDING_MODEL: ${EMBEDDING_MODEL:-text-embedding-3-small}
      EMBEDDING_DIM: ${EMBEDDING_DIM:-1536}
      RATE_LIMIT_PER_MIN: ${RATE_LIMIT_PER_MIN:-150}
      NEO4J_URI: bolt://neo4j:7687
      NEO4J_USER: neo4j
      NEO4J_PASSWORD: password
      REDIS_URL: redis://redis:6379/0
      POSTGRES_DSN: postgresql://cogram:cogram@postgres:5432/cogram
      TRAINER_URL: http://cogram-trainer:7900
      MCP_HTTP_PORT: 7800
      MAX_DAILY_INGEST_CALLS: ${MAX_DAILY_INGEST_CALLS:-300}
      COGRAM_FULL_PIPELINE: "true"
    volumes:
      - cogram_episodes:/app/episodes
      - cogram_cache:/app/cache
    depends_on:
      neo4j:
        condition: service_healthy
      redis:
        condition: service_healthy
      postgres:
        condition: service_healthy
    healthcheck:
      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:7800/health')"]
      interval: 60s
      timeout: 5s
      retries: 5
      start_period: 20s
    networks:
      - cogram

  # ==========================================================================
  # Cogram trainer (Qwen / Gemma + PEFT + LoRA scheduler)
  # GPU optional; falls back to CPU. Uses VRAM coexistence (5-layer).
  #
  # Purpose: trains per-node LoRA adapters from accumulated synthesis pairs in
  # engram.training_data. Once trained, those adapters REPLACE Gemma for that
  # node's narrations — smaller, faster, personalized to the user's voice.
  #
  # Opt-in only: requires `--profile training` to start, since it needs ≥50
  # samples per node before doing anything useful (months of normal usage).
  # `docker compose up` skips trainer; `docker compose --profile training up`
  # brings it in.
  # ==========================================================================
  cogram-trainer:
    profiles: ["training"]
    build:
      context: .
      dockerfile: containers/cogram-trainer/Dockerfile
    container_name: cogram-trainer
    ports:
      - "7900:7900"
    environment:
      POSTGRES_DSN: postgresql://cogram:cogram@postgres:5432/cogram
      BASE_MODEL: ${BASE_MODEL:-Qwen/Qwen2.5-1.5B-Instruct}
      TRAINER_VRAM_FRACTION: ${TRAINER_VRAM_FRACTION:-0.40}
      TRAINER_GPU_UTIL_PAUSE: ${TRAINER_GPU_UTIL_PAUSE:-50}
      TRAINER_SCHEDULE: ${TRAINER_SCHEDULE:-02:00-06:00}
      TRAINER_REQUIRE_USER_IDLE: ${TRAINER_REQUIRE_USER_IDLE:-false}
      TRAINER_MIN_SAMPLES: ${TRAINER_MIN_SAMPLES:-20}
      TRAINER_MIN_DAYS_BETWEEN: ${TRAINER_MIN_DAYS_BETWEEN:-7}
      HF_HOME: /app/hf_cache
    volumes:
      - cogram_adapters:/cogram/adapters
      - cogram_hf_cache:/app/hf_cache
    deploy:
      resources:
        limits:
          memory: 8G
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    depends_on:
      postgres:
        condition: service_healthy
    networks:
      - cogram
    # If you have no NVIDIA GPU, comment out the deploy.resources block above.

  # ==========================================================================
  # Cogram dashboard (FastAPI + force-graph viz + live metrics on :7801)
  # Image: pulled from ghcr.io by default. Falls back to local build.
  # ==========================================================================
  cogram-dashboard:
    image: ghcr.io/srk0102/cogram-dashboard:latest
    pull_policy: missing
    build:
      context: .
      dockerfile: containers/cogram-dashboard/Dockerfile
    container_name: cogram-dashboard
    ports:
      - "7801:7801"
    environment:
      NEO4J_URI: bolt://neo4j:7687
      NEO4J_USER: neo4j
      NEO4J_PASSWORD: password
      REDIS_URL: redis://redis:6379/0
      POSTGRES_DSN: postgresql://cogram:cogram@postgres:5432/cogram
      MCP_URL: http://cogram-mcp:7800
      DASHBOARD_PORT: 7801
    depends_on:
      neo4j:
        condition: service_healthy
      redis:
        condition: service_healthy
      postgres:
        condition: service_healthy
    networks:
      - cogram

volumes:
  neo4j_data:
  neo4j_logs:
  redis_data:
  postgres_data:
  cogram_episodes:
  cogram_cache:
  cogram_adapters:
  cogram_hf_cache:

networks:
  cogram:
    driver: bridge