LightRAG/docker-compose.test.yml

name: lightrag-entity-resolution-test

services:
  postgres:
    container_name: lightrag-postgres
    build:
      context: ./docker/postgres-age-vector
      dockerfile: Dockerfile
    environment:
      POSTGRES_DB: lightrag
      POSTGRES_USER: lightrag
      POSTGRES_PASSWORD: lightrag_pass
    ports:
      - "5433:5432"  # Use 5433 to avoid conflict with agent-sdk postgres
    volumes:
      - pgdata_test:/var/lib/postgresql/data
    command: |
      postgres
      -c shared_preload_libraries='vector,age'
      -c max_connections=150
      -c shared_buffers=768MB
      -c work_mem=32MB
      -c checkpoint_completion_target=0.9
      -c effective_cache_size=2GB
      -c maintenance_work_mem=192MB
      -c wal_compression=on
      -c checkpoint_timeout=10min
      -c max_wal_size=1GB
      -c random_page_cost=1.1
      -c effective_io_concurrency=200
      -c max_worker_processes=12
      -c max_parallel_workers_per_gather=4
      -c max_parallel_workers=8
      -c max_parallel_maintenance_workers=4
      -c jit_above_cost=50000
      -c jit_inline_above_cost=250000
      -c jit_optimize_above_cost=250000
      -c default_statistics_target=200
      -c hash_mem_multiplier=4
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U lightrag -d lightrag"]
      interval: 5s
      timeout: 5s
      retries: 5
    mem_limit: 2g

  lightrag:
    container_name: lightrag-test
    build:
      context: .
      dockerfile: Dockerfile
    ports:
      - "9622:9621"  # Use 9622 to avoid conflict
    volumes:
      - ./data/rag_storage_test:/app/data/rag_storage
      - ./data/inputs_test:/app/data/inputs
      # Live reload: Use absolute host path for Docker-in-Docker compatibility (Coder)
      - /var/lib/docker/volumes/coder-shared-projects-optimized/_data/LightRAG/lightrag:/app/lightrag
    environment:
      # Live reload: PYTHONPATH makes mounted /app/lightrag take precedence over site-packages
      - PYTHONPATH=/app

      # Server
      - HOST=0.0.0.0
      - PORT=9621
      - LOG_LEVEL=DEBUG

      # LLM (OpenAI - gpt-4o-mini for reliable fast extraction)
      - LLM_BINDING=openai
      - LLM_MODEL=gpt-4o-mini
      - LLM_BINDING_HOST=https://api.openai.com/v1
      - LLM_BINDING_API_KEY=${OPENAI_API_KEY}

      # Embedding
      - EMBEDDING_BINDING=openai
      - EMBEDDING_MODEL=text-embedding-3-small
      - EMBEDDING_DIM=1536
      - EMBEDDING_BINDING_HOST=https://api.openai.com/v1
      - EMBEDDING_BINDING_API_KEY=${OPENAI_API_KEY}

      # Storage Configuration - Full PostgreSQL!
      # Custom postgres image has pgvector + Apache AGE
      - LIGHTRAG_KV_STORAGE=PGKVStorage
      - LIGHTRAG_VECTOR_STORAGE=PGVectorStorage
      - LIGHTRAG_GRAPH_STORAGE=PGGraphStorage
      - LIGHTRAG_DOC_STATUS_STORAGE=PGDocStatusStorage
      - POSTGRES_HOST=postgres
      - POSTGRES_PORT=5432
      - POSTGRES_USER=lightrag
      - POSTGRES_PASSWORD=lightrag_pass
      - POSTGRES_DATABASE=lightrag

      # Entity Resolution - DISABLED for faster ingestion (testing Context Precision changes)
      - ENTITY_RESOLUTION_ENABLED=false
      - ENTITY_RESOLUTION_FUZZY_THRESHOLD=0.85
      - ENTITY_RESOLUTION_VECTOR_THRESHOLD=0.5
      - ENTITY_RESOLUTION_MAX_CANDIDATES=3

      # Orphan Connection - MANUAL (use UI button instead of auto)
      - AUTO_CONNECT_ORPHANS=false

      # Processing - Matching agent-sdk working settings
      - MAX_ASYNC=96
      - MAX_PARALLEL_INSERT=10
      - EMBEDDING_FUNC_MAX_ASYNC=2             # Match llamacpp parallel slots (prevent queue backlog)
      - EMBEDDING_BATCH_NUM=48

      # Gunicorn - 8 workers x 4 threads = 32 concurrent handlers
      - GUNICORN_CMD_ARGS=--workers=8 --worker-class=gthread --threads=4 --worker-connections=1000 --timeout=120 --keep-alive=5 --graceful-timeout=30

      # Extraction - Using agent-sdk defaults for reliable ingestion
      - CHUNK_SIZE=1200                        # Default chunk size (agent-sdk default)
      - CHUNK_OVERLAP_SIZE=100                 # Default overlap
      # MAX_GLEANING defaults to 1 (removed override of 2)

      # Orphan Connection - Use UI button for manual triggering
      # AUTO_CONNECT_ORPHANS is set to false above (manual mode)
      - ORPHAN_CONNECTION_THRESHOLD=0.3        # Vector similarity pre-filter threshold
      - ORPHAN_CONFIDENCE_THRESHOLD=0.7        # LLM confidence required for connection
      - ORPHAN_CROSS_CONNECT=true              # Allow orphan-to-orphan connections
    depends_on:
      postgres:
        condition: service_healthy
    entrypoint: []
    command:
      - python
      - -m
      - lightrag.api.run_with_gunicorn
      - --workers
      - "8"
      - --llm-binding
      - openai
      - --embedding-binding
      - openai
    healthcheck:
      test: ["CMD-SHELL", "curl -f http://localhost:9621/health || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 10
      start_period: 60s
    mem_limit: 2g

volumes:
  pgdata_test: