didi-lot1-ai/ai_platform/modules/embeddings/deploy/docker-compose.yml

# Embeddings Module - Docker Compose Configuration
#
# Port Allocation (x41xx = Embeddings):
#   14100 - Embeddings API (Prod)
#   54100 - Embeddings API (Dev)
#   14101/54101 - vLLM Embedding Server
#   14110/54110 - llama.cpp Embedding Server
#
# Profiles:
#   api      - API server only (uses external embedding servers)
#   vllm     - API + vLLM server (GPU required)
#   llamacpp - API + llama.cpp server
#
# Naming Convention: didiAI-{module}-{service}
#
# Network:
#   Uses didi-network (shared with all DIDI + AI platform stacks)

networks:
  didi-network:
    external: true

services:
  # ==========================================================================
  # Embeddings API Server
  # ==========================================================================
  embeddings-api:
    container_name: didiAI-embeddings-api
    image: didiai-embeddings-api
    build:
      context: ..
      dockerfile: deploy/Dockerfile
    ports:
      - "${EMB_PORT:-14100}:${EMB_PORT:-14100}"
    networks:
      - didi-network
    environment:
      - EMB_PORT=${EMB_PORT:-14100}
      - EMB_EXTERNAL_URL=${EMB_EXTERNAL_URL}
      - EMB_DEFAULT_BACKEND=${EMB_DEFAULT_BACKEND}
      - EMB_ENABLE_VLLM=${EMB_ENABLE_VLLM}
      - EMB_ENABLE_LLAMACPP=${EMB_ENABLE_LLAMACPP}
      - EMB_VLLM_BASE_URL=http://didiAI-embeddings-vllm:14101
      - EMB_LLAMACPP_BASE_URL=http://didiAI-embeddings-llamacpp:8080
      - EMB_API_TOKENS=${EMB_API_TOKENS:-}
      - EMB_DASHBOARD_URL=${EMB_DASHBOARD_URL:-http://didiAI-dashboard:51300}
    healthcheck:
      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:${EMB_PORT:-14100}/health')"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 30s
    restart: unless-stopped
    profiles:
      - api
      - vllm
      - llamacpp

  # ==========================================================================
  # vLLM Embedding Server
  # ==========================================================================
  # Embedding model using vLLM with --task embed (v0.8.x syntax)
  vllm-embed:
    container_name: didiAI-embeddings-vllm
    image: vllm/vllm-openai:v0.8.5
    ports:
      - "${EMB_VLLM_PORT:-14101}:14101"
    networks:
      - didi-network
    volumes:
      - ${HF_CACHE_DIR:-/cai2_ds_storage/hf_cache}:/root/.cache/huggingface
    environment:
      - HF_HOME=/root/.cache/huggingface
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
      - CUDA_VISIBLE_DEVICES=${EMB_VLLM_GPU:-0}
    command: >
      --model ${EMB_VLLM_MODEL:-BAAI/bge-m3}
      --host 0.0.0.0
      --port 14101
      --task embed
      --trust-remote-code
      --max-model-len ${EMB_VLLM_MAX_LEN:-8192}
      --gpu-memory-utilization ${EMB_VLLM_GPU_UTIL:-0.50}
      --disable-log-requests
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['${EMB_VLLM_GPU:-0}']
              capabilities: [gpu]
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:14101/health"]
      interval: 30s
      timeout: 10s
      retries: 10
      start_period: 300s
    restart: unless-stopped
    profiles:
      - vllm

  # ==========================================================================
  # llama.cpp Embedding Server
  # ==========================================================================
  # Embedding model using llama.cpp with --embedding
  llamacpp-embed:
    container_name: didiAI-embeddings-llamacpp
    image: ghcr.io/ggml-org/llama.cpp:server-b4769
    ports:
      - "${EMB_LLAMACPP_PORT:-14110}:8080"
    networks:
      - didi-network
    volumes:
      - ${MODELS_DIR:-/cai2_ds_storage/models}:/models:ro
    command: >
      --model /models/${EMB_LLAMACPP_MODEL:-bge-m3-q4_k_m.gguf}
      --host 0.0.0.0
      --port 8080
      --embedding
      --ctx-size ${EMB_LLAMACPP_CTX:-8192}
      --threads ${EMB_LLAMACPP_THREADS:-4}
      --parallel ${EMB_LLAMACPP_PARALLEL:-4}
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8080/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 60s
    restart: unless-stopped
    profiles:
      - llamacpp