docker-compose.yml

name: interview-agent-python

services:
  mysql:
    image: ${MYSQL_IMAGE:-mysql:8.4}
    environment:
      MYSQL_ROOT_PASSWORD: interview-root
      MYSQL_DATABASE: interview_agent
      MYSQL_USER: interview
      MYSQL_PASSWORD: interview
      TZ: Asia/Shanghai
    command:
      - --character-set-server=utf8mb4
      - --collation-server=utf8mb4_unicode_ci
    ports:
      - "${MYSQL_PORT:-3307}:3306"
    volumes:
      - mysql_data:/var/lib/mysql
    healthcheck:
      test: ["CMD", "mysqladmin", "ping", "-h", "127.0.0.1", "-uinterview", "-pinterview"]
      interval: 10s
      timeout: 5s
      retries: 15

  redis:
    image: redis:7.4-alpine
    command: ["redis-server", "--appendonly", "yes", "--appendfsync", "everysec"]
    ports:
      - "${REDIS_PORT:-6381}:6379"
    volumes:
      - redis_data:/data
    healthcheck:
      test: ["CMD", "redis-cli", "ping"]
      interval: 5s
      timeout: 3s
      retries: 15

  etcd:
    image: quay.io/coreos/etcd:v3.5.18
    environment:
      ETCD_AUTO_COMPACTION_MODE: revision
      ETCD_AUTO_COMPACTION_RETENTION: "1000"
      ETCD_QUOTA_BACKEND_BYTES: "4294967296"
      ETCD_SNAPSHOT_COUNT: "50000"
    command: >-
      etcd --advertise-client-urls=http://127.0.0.1:2379
      --listen-client-urls=http://0.0.0.0:2379 --data-dir=/etcd
    volumes:
      - etcd_data:/etcd
    healthcheck:
      test: ["CMD", "etcdctl", "endpoint", "health"]
      interval: 10s
      timeout: 5s
      retries: 15

  minio:
    image: minio/minio:RELEASE.2024-12-18T13-15-44Z
    environment:
      MINIO_ROOT_USER: minioadmin
      MINIO_ROOT_PASSWORD: minioadmin
    command: minio server /data --console-address :9001
    volumes:
      - minio_data:/data
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:9000/minio/health/live"]
      interval: 10s
      timeout: 5s
      retries: 15

  milvus:
    image: milvusdb/milvus:v2.6.4
    command: ["milvus", "run", "standalone"]
    security_opt:
      - seccomp:unconfined
    environment:
      ETCD_ENDPOINTS: etcd:2379
      MINIO_ADDRESS: minio:9000
    ports:
      - "${MILVUS_PORT:-19531}:19530"
      - "${MILVUS_HEALTH_PORT:-9093}:9091"
    volumes:
      - milvus_data:/var/lib/milvus
    depends_on:
      etcd:
        condition: service_healthy
      minio:
        condition: service_healthy
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:9091/healthz"]
      interval: 10s
      timeout: 5s
      retries: 30

  embeddings:
    # TEI currently publishes numbered x86_64 tags but only a rolling ARM64 tag.
    # Pin the resolved ARM64 index digest; x86_64 users override TEI_IMAGE as documented.
    image: ${TEI_IMAGE:-ghcr.io/huggingface/text-embeddings-inference@sha256:35c50d7494de22deecdb783b8f5b7e1d05765709bd90071b03469b9440d28656}
    # BGE-M3 advertises an 8k context. TEI's default 16k-token warm-up batch can
    # exceed Docker Desktop's memory limit on CPU/ARM64, so bound dynamic batches
    # to the question-bank chunk size used by this service.
    command:
      - --model-id
      - BAAI/bge-m3
      - --port
      - "80"
      - --max-batch-tokens
      - "1024"
      - --max-batch-requests
      - "8"
      - --max-concurrent-requests
      - "32"
    ports:
      - "8081:80"
    volumes:
      - hf_cache:/data
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost/health"]
      interval: 20s
      timeout: 10s
      retries: 60
      start_period: 60s

  reranker-models:
    # Install a pinned INT8 ONNX export so the multilingual reranker can coexist
    # with BGE-M3 under an 8 GiB Docker Desktop memory limit.
    image: ${TEI_IMAGE:-ghcr.io/huggingface/text-embeddings-inference@sha256:35c50d7494de22deecdb783b8f5b7e1d05765709bd90071b03469b9440d28656}
    entrypoint: ["/bin/sh", "-ec"]
    command:
      - >-
        base=https://huggingface.co/cross-encoder/mmarco-mMiniLMv2-L12-H384-v1/resolve/1427fd652930e4ba29e8149678df786c240d8825;
        mkdir -p /models/onnx;
        for file in config.json tokenizer.json tokenizer_config.json special_tokens_map.json; do
          test -s /models/$${file} || curl -fL --retry 5 -o /models/$${file} $${base}/$${file};
        done;
        test -s /models/onnx/model.onnx || curl -fL --retry 5 -o /models/onnx/model.onnx $${base}/onnx/${RERANK_ONNX_FILE:-model_qint8_arm64.onnx};
        test -s /models/onnx/model.onnx
    volumes:
      - reranker_models:/models
    restart: "no"

  reranker:
    # Reuse the TEI runtime image; weights come from the pinned installer above.
    image: ${TEI_IMAGE:-ghcr.io/huggingface/text-embeddings-inference@sha256:35c50d7494de22deecdb783b8f5b7e1d05765709bd90071b03469b9440d28656}
    command:
      - --model-id
      - /models
      - --port
      - "80"
      - --tokenization-workers
      - "2"
      - --max-batch-tokens
      - "512"
      - --max-batch-requests
      - "1"
      - --max-concurrent-requests
      - "4"
    ports:
      - "${RERANK_PORT:-8082}:80"
    volumes:
      - reranker_models:/models:ro
    depends_on:
      reranker-models:
        condition: service_completed_successfully
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost/health"]
      interval: 20s
      timeout: 10s
      retries: 60
      start_period: 60s

  speech:
    # One ARM64/AMD64 CPU image provides both local Whisper ASR and Kokoro TTS.
    image: ${SPEACHES_IMAGE:-ghcr.io/speaches-ai/speaches@sha256:21e3df06d842fb7802ab470dd77c25f0e8c0d22950e8d8c6ae886e851af53ef8}
    environment:
      WHISPER__COMPUTE_TYPE: int8
      WHISPER__CPU_THREADS: "4"
      WHISPER__TTL: "60"
    ports:
      - "${SPEECH_PORT:-8002}:8000"
    volumes:
      # Speaches runs as a non-root user; use its own persistent cache so the
      # root-owned TEI cache cannot cause permission errors.
      - speech_hf_cache:/home/ubuntu/.cache/huggingface/hub
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 15s
      timeout: 10s
      retries: 60
      start_period: 300s

  speech-models:
    # Reuse the exact Speaches image as a one-shot ASR model installer. Repeated
    # starts are fast because the Whisper snapshot already exists in the volume.
    image: ${SPEACHES_IMAGE:-ghcr.io/speaches-ai/speaches@sha256:21e3df06d842fb7802ab470dd77c25f0e8c0d22950e8d8c6ae886e851af53ef8}
    entrypoint: ["/bin/sh", "-ec"]
    command:
      - >-
        curl -fsS --retry 5 -X POST
        http://speech:8000/v1/models/${SPEECH_TRANSCRIPTION_MODEL:-Systran/faster-whisper-base}
    depends_on:
      speech:
        condition: service_healthy
    restart: "no"

  tts-models:
    build: .
    entrypoint: ["/bin/sh", "-ec"]
    command:
      - >-
        if [ ! -s /models/sherpa-onnx-vits-zh-ll/model.onnx ]; then
          curl -fL --retry 5 -o /models/sherpa-onnx-vits-zh-ll.tar.bz2 https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/sherpa-onnx-vits-zh-ll.tar.bz2;
          tar -xjf /models/sherpa-onnx-vits-zh-ll.tar.bz2 -C /models;
          rm /models/sherpa-onnx-vits-zh-ll.tar.bz2;
        fi;
        test -s /models/sherpa-onnx-vits-zh-ll/model.onnx;
        test -s /models/sherpa-onnx-vits-zh-ll/lexicon.txt;
        test -s /models/sherpa-onnx-vits-zh-ll/tokens.txt
    volumes:
      - tts_models:/models
    restart: "no"

  app:
    profiles: ["app"]
    build: .
    env_file: .env
    environment:
      MYSQL_DSN: mysql+asyncmy://interview:interview@mysql:3306/interview_agent?charset=utf8mb4
      REDIS_URL: redis://redis:6379/0
      MILVUS_URI: http://milvus:19530
      EMBEDDING_BASE_URL: http://embeddings:80
      RERANK_BASE_URL: http://reranker:80
      SPEECH_BASE_URL: http://speech:8000
      TTS_MODEL_DIRECTORY: /models/sherpa-onnx-vits-zh-ll
    ports:
      - "${APP_PORT:-9092}:9090"
    depends_on:
      mysql:
        condition: service_healthy
      redis:
        condition: service_healthy
      milvus:
        condition: service_healthy
      embeddings:
        condition: service_healthy
      reranker:
        condition: service_healthy
      speech:
        condition: service_healthy
      speech-models:
        condition: service_completed_successfully
      tts-models:
        condition: service_completed_successfully
    volumes:
      - tts_models:/models:ro
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:9090/health/ready"]
      interval: 10s
      timeout: 5s
      retries: 20
      start_period: 20s

volumes:
  mysql_data:
  redis_data:
  etcd_data:
  minio_data:
  milvus_data:
  hf_cache:
  speech_hf_cache:
  tts_models:
  reranker_models: