WeLe Agentic AI with Docker deployment

docker-compose.yml

@@ -28,6 +28,12 @@ services:
       - 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
@@ -50,13 +56,22 @@ services:
     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: 768M
+          memory: 1024M
 
     logging:
       driver: json-file
@@ -75,3 +90,4 @@ networks:
 
 volumes:
   artifacts:
+  speech_cache: