# ============================================
# WeLe Agentic AI — deployment
#
# Runs alongside the existing CRM stack rather than inside it. Joining the
# CRM's network (`whatsapp-crm_default`) lets this service reach `redis` and
# `whatsapp-api` by name, without publishing anything extra or duplicating a
# Redis on a 3.7 GB box.
#
# The voice service is not here on purpose — it needs a CUDA GPU. See DEPLOY.md.
# ============================================

services:
  agentic-ai:
    build: .
    image: wele/agentic-ai:latest
    container_name: wele-agentic-ai
    restart: always

    # Host 4000 is free. The CRM's chat-service also listens on 4000, but only
    # inside its own container — it publishes nothing, so there is no clash.
    ports:
      - "4000:4000"

    env_file:
      - .env

    environment:
      - NODE_ENV=production
      - PORT=4000

      # English-only keeps one TTS voice resident (~200 MB). Tamil is built and
      # tested — VOICE_LANGUAGES=en,ta plus a multilingual STT_MODEL enables it,
      # at roughly 200 MB more.
      - VOICE_LANGUAGES=en
      - STT_MODEL=onnx-community/whisper-tiny.en

      # Reuse the CRM's Redis by service name on the shared network.
      # Keys are namespaced with REDIS_PREFIX, so the two never collide.
      - REDIS_ENABLED=true
      - REDIS_HOST=redis
      - REDIS_PORT=6379
      # Quoted: a trailing colon makes YAML parse this as a map, not a string.
      - "REDIS_PREFIX=agentic:"

      # Writes go through the CRM's own REST routes so its lead scoring,
      # socket events and audit trail still fire. Reached over the shared
      # network, so this never leaves the host.
      - "CRM_API_BASE=http://whatsapp-api:3000"

      # Must match the CRM's JWT_SECRET — that is how a CRM login is accepted
      # here with the same role and permissions. Sourced from .env so the
      # secret is never written into this file.
      - CRM_JWT_SECRET=${CRM_JWT_SECRET:?set CRM_JWT_SECRET in .env — must equal the CRM JWT_SECRET}
      - MONGODB_URI=${MONGODB_URI:?set MONGODB_URI in .env}

    volumes:
      # Generated xlsx/pdf/pptx survive rebuilds; swept on a TTL by the app.
      - artifacts:/app/storage/artifacts
      # Speech models are fetched from HuggingFace on first use (~200 MB).
      # Without this they re-download on every restart and the first voice turn
      # after a deploy stalls for a minute.
      - speech_cache:/app/.transformers-cache

    # A 2 vCPU / 3.7 GB host already runs the CRM, chat-service, Redis and
    # Milvus. Capping this container keeps a runaway turn from starving them.
    #
    # 1 GB, not 768 MB: the speech models are resident once voice is used —
    # measured 729 MB (Whisper tiny.en 415 MB + MMS-TTS English 203 MB + VAD
    # 28 MB + the agent itself). 768 MB left no headroom, and an OOM kill takes
    # text chat down with voice. Text-only sessions stay near 80 MB.
    deploy:
      resources:
        limits:
          memory: 1024M

    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"

    networks:
      - crm

networks:
  crm:
    # Created by the CRM's own compose project; we attach, never own it.
    name: whatsapp-crm_default
    external: true

volumes:
  artifacts:
  speech_cache: