Voice Assistant

End-to-end voice assistant pipeline for the DaviesTechLabs AI/ML platform.

Components

Real-time Handler (NATS-based)

The voice assistant service listens on NATS for audio requests and returns synthesized speech responses. It uses the handler-base library for standardized NATS handling, telemetry, and health checks.

Pipeline: STT → Embeddings → Milvus RAG → Rerank → LLM → TTS

Kubeflow Pipeline (Batch)

For batch processing or async workflows via Kubeflow Pipelines.

Pipeline	Description
voice_pipeline.yaml	Full STT → RAG → TTS pipeline
rag_pipeline.yaml	Text-only RAG pipeline
tts_pipeline.yaml	Simple text-to-speech

# Compile pipelines
cd pipelines
pip install kfp==2.12.1
python voice_pipeline.py

Architecture

NATS (voice.request)
        │
        ▼
┌───────────────────┐
│  Voice Assistant  │
│    Handler        │
└───────────────────┘
        │
        ├──▶ Whisper STT (elminster)
        │         │
        │         ▼
        ├──▶ BGE Embeddings (drizzt)
        │         │
        │         ▼
        ├──▶ Milvus Vector Search
        │         │
        │         ▼
        ├──▶ BGE Reranker (danilo)
        │         │
        │         ▼
        ├──▶ vLLM (khelben)
        │         │
        │         ▼
        └──▶ XTTS TTS (elminster)
                  │
                  ▼
         NATS (voice.response.{id})

Configuration

Environment Variable	Default	Description
NATS_URL	nats://nats.ai-ml.svc.cluster.local:4222	NATS server
WHISPER_URL	http://whisper-predictor.ai-ml.svc.cluster.local	STT service
EMBEDDINGS_URL	http://embeddings-predictor.ai-ml.svc.cluster.local	Embeddings
RERANKER_URL	http://reranker-predictor.ai-ml.svc.cluster.local	Reranker
VLLM_URL	http://llm-draft.ai-ml.svc.cluster.local:8000	LLM service
TTS_URL	http://tts-predictor.ai-ml.svc.cluster.local	TTS service
MILVUS_HOST	milvus.ai-ml.svc.cluster.local	Vector DB
COLLECTION_NAME	knowledge_base	Milvus collection

NATS Message Format

Request (voice.request)

{
  "request_id": "uuid",
  "audio": "base64-encoded-audio",
  "language": "en",
  "collection": "knowledge_base"
}

Response (voice.response.{request_id})

{
  "request_id": "uuid",
  "transcription": "user question",
  "response": "assistant answer",
  "audio": "base64-encoded-audio"
}

Building

docker build -t voice-assistant:latest .

# With specific handler-base tag
docker build --build-arg BASE_TAG=latest -t voice-assistant:latest .

Related

• homelab-design - Architecture docs
• kuberay-images - Ray worker images
• handler-base - Base handler library