docker-compose.yml

x-ai-worker-common: &ai-worker-common
  build:
    context: .
    dockerfile: ai_worker/Dockerfile
    args:
      MODEL_SOURCE: ${MODEL_SOURCE:-remote}
      YOLO_MODEL_NAME: ${YOLO_MODEL_NAME:-yolo11n}
      YOLO_MODEL_RELEASE: ${YOLO_MODEL_RELEASE:-v8.3.0}
      YOLO_MODEL_URL: ${YOLO_MODEL_URL:-}
  image: rtspanda-ai-worker:latest
  environment:
    MODEL_SOURCE: "${MODEL_SOURCE:-remote}"
    MODEL_PATH: "${MODEL_PATH:-/model/model.onnx}"
    YOLO_MODEL_PATH: "${MODEL_PATH:-/model/model.onnx}"
    YOLO_MODEL_URL: "${YOLO_MODEL_URL:-}"
    YOLO_CONFIDENCE: "${YOLO_CONFIDENCE:-0.25}"
    YOLO_IOU: "${YOLO_IOU:-0.45}"
    YOLO_MAX_DETECTIONS: "${YOLO_MAX_DETECTIONS:-100}"
  healthcheck:
    test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8090/health', timeout=3)"]
    interval: 30s
    timeout: 5s
    retries: 5
    start_period: 20s
  restart: unless-stopped
  mem_limit: 512m
  deploy:
    resources:
      limits:
        memory: 512m

x-rtspanda-common: &rtspanda-common
  build:
    context: .
    args:
      MEDIAMTX_VERSION: ${MEDIAMTX_VERSION:-v1.12.3}
  image: rtspanda:latest
  ports:
    - "8080:8080"
  environment:
    DATA_DIR: /data
    PORT: "8080"
    MEDIAMTX_BIN: /usr/local/bin/mediamtx
    FFMPEG_BIN: /usr/bin/ffmpeg
    AUTH_ENABLED: "${AUTH_ENABLED:-false}"
    AUTH_TOKEN: "${AUTH_TOKEN:-}"
    MEDIAMTX_SOURCE_ON_DEMAND: "${MEDIAMTX_SOURCE_ON_DEMAND:-true}"
    MEDIAMTX_SOURCE_ON_DEMAND_CLOSE_AFTER: "${MEDIAMTX_SOURCE_ON_DEMAND_CLOSE_AFTER:-10s}"
    MEDIAMTX_HLS_ALWAYS_REMUX: "${MEDIAMTX_HLS_ALWAYS_REMUX:-false}"
    MEDIAMTX_HLS_SEGMENT_COUNT: "${MEDIAMTX_HLS_SEGMENT_COUNT:-3}"
    MEDIAMTX_HLS_SEGMENT_DURATION: "${MEDIAMTX_HLS_SEGMENT_DURATION:-2s}"
    MEDIAMTX_HLS_PART_DURATION: "${MEDIAMTX_HLS_PART_DURATION:-200ms}"
    # fmp4 = standard HLS (hls.js); lowLatency can 404 index.m3u8 for some RTSP cameras.
    MEDIAMTX_HLS_VARIANT: "${MEDIAMTX_HLS_VARIANT:-fmp4}"
    MEDIAMTX_LOG_LEVEL: "${MEDIAMTX_LOG_LEVEL:-warn}"
    MEDIAMTX_RTSP_TRANSPORT: "${MEDIAMTX_RTSP_TRANSPORT:-tcp}"
    RTSPANDA_MODE: "${RTSPANDA_MODE:-standard}"
    AI_MODE: "${AI_MODE:-local}"
    AI_WORKER_URL: "${AI_WORKER_URL:-}"
    DETECTOR_URL: "${DETECTOR_URL:-}"
    DETECTION_SAMPLE_INTERVAL_SECONDS: "${DETECTION_SAMPLE_INTERVAL_SECONDS:-30}"
    DETECTION_WORKERS: "${DETECTION_WORKERS:-1}"
    DETECTION_QUEUE_SIZE: "${DETECTION_QUEUE_SIZE:-32}"
    GOMEMLIMIT: "${GOMEMLIMIT:-200MiB}"
  volumes:
    - ./data:/data
  healthcheck:
    test: ["CMD", "wget", "-qO-", "http://127.0.0.1:8080/api/v1/health"]
    interval: 30s
    timeout: 5s
    retries: 5
    start_period: 20s
  restart: unless-stopped
  mem_limit: 512m
  deploy:
    resources:
      limits:
        memory: 512m

services:
  # Default local stack. This preserves the existing `docker compose up --build -d`
  # workflow for x86/GPU users and for single-machine development.
  ai-worker:
    <<: *ai-worker-common
    container_name: rtspanda-ai-worker
    expose:
      - "8090"

  rtspanda:
    <<: *rtspanda-common
    container_name: rtspanda
    depends_on:
      ai-worker:
        condition: service_healthy
        required: false

  # Pi-first lightweight mode: backend + mediamtx only. Detection is disabled
  # until AI_WORKER_URL (or DETECTOR_URL) points at a remote AI worker.
  rtspanda-pi:
    <<: *rtspanda-common
    container_name: rtspanda-pi
    profiles: ["pi"]
    environment:
      DATA_DIR: /data
      PORT: "8080"
      MEDIAMTX_BIN: /usr/local/bin/mediamtx
      FFMPEG_BIN: /usr/bin/ffmpeg
      AUTH_ENABLED: "${AUTH_ENABLED:-false}"
      AUTH_TOKEN: "${AUTH_TOKEN:-}"
      MEDIAMTX_SOURCE_ON_DEMAND: "${MEDIAMTX_SOURCE_ON_DEMAND:-true}"
      MEDIAMTX_SOURCE_ON_DEMAND_CLOSE_AFTER: "${MEDIAMTX_SOURCE_ON_DEMAND_CLOSE_AFTER:-10s}"
      MEDIAMTX_HLS_ALWAYS_REMUX: "${MEDIAMTX_HLS_ALWAYS_REMUX:-false}"
      MEDIAMTX_HLS_SEGMENT_COUNT: "${MEDIAMTX_HLS_SEGMENT_COUNT:-3}"
      MEDIAMTX_HLS_SEGMENT_DURATION: "${MEDIAMTX_HLS_SEGMENT_DURATION:-2s}"
      MEDIAMTX_HLS_PART_DURATION: "${MEDIAMTX_HLS_PART_DURATION:-200ms}"
      MEDIAMTX_HLS_VARIANT: "${MEDIAMTX_HLS_VARIANT:-fmp4}"
      MEDIAMTX_LOG_LEVEL: "${MEDIAMTX_LOG_LEVEL:-warn}"
      MEDIAMTX_RTSP_TRANSPORT: "${MEDIAMTX_RTSP_TRANSPORT:-tcp}"
      # Pi mode: viewer + stream relay + snapshot AI only.
      # Real-time YOLO inference is NOT supported on Raspberry Pi.
      RTSPANDA_MODE: "pi"
      AI_MODE: "remote"
      AI_WORKER_URL: "${AI_WORKER_URL:-}"
      DETECTOR_URL: "${DETECTOR_URL:-}"
      DETECTION_SAMPLE_INTERVAL_SECONDS: "${DETECTION_SAMPLE_INTERVAL_SECONDS:-30}"
      DETECTION_WORKERS: "${DETECTION_WORKERS:-1}"
      DETECTION_QUEUE_SIZE: "${DETECTION_QUEUE_SIZE:-32}"
      GOMEMLIMIT: "${GOMEMLIMIT:-200MiB}"
      # Snapshot Intelligence Engine (Pi AI replacement)
      SNAPSHOT_AI_ENABLED: "${SNAPSHOT_AI_ENABLED:-false}"
      SNAPSHOT_AI_PROVIDER: "${SNAPSHOT_AI_PROVIDER:-claude}"
      SNAPSHOT_AI_API_KEY: "${SNAPSHOT_AI_API_KEY:-}"
      SNAPSHOT_AI_INTERVAL_SECONDS: "${SNAPSHOT_AI_INTERVAL_SECONDS:-30}"
      SNAPSHOT_AI_PROMPT: "${SNAPSHOT_AI_PROMPT:-Detect people, vehicles, packages, or notable activity near a property.}"
      SNAPSHOT_AI_THRESHOLD: "${SNAPSHOT_AI_THRESHOLD:-medium}"

  # Standalone AI-worker profile for a second machine or dedicated inference host.
  ai-worker-standalone:
    <<: *ai-worker-common
    container_name: rtspanda-ai-worker-standalone
    profiles: ["ai-worker"]
    ports:
      - "8090:8090"

  # Explicit full profile alias for environments that want a named profile
  # without changing the default unprofiled stack behavior.
  stack-full:
    image: alpine:3.22
    profiles: ["full"]
    command: ["sh", "-c", "tail -f /dev/null"]
    depends_on:
      ai-worker:
        condition: service_healthy
      rtspanda:
        condition: service_healthy