Vaani Sahayak (वाणी सहायक) — Sovereign AI Assistant for Government Welfare Schemes
Vaani Sahayak is a sovereign AI assistant that helps citizens discover and understand Indian
government welfare schemes. Powered entirely by Indian-built AI models, with no dependency on
OpenAI or any external cloud API.
Developed as an open-source blueprint under the Cloud2 Labs Innovation Hub, Vaani Sahayak
demonstrates how India's homegrown language models, retrieval-augmented generation, and on-
device Indic TTS can be combined into a practical, fully offline citizen-facing assistant. It indexes
2,086 schemes from myscheme.gov.in across agriculture, health, education, housing, social welfare,
and more.
It showcases a production-style microservices architecture: a FastAPI backend for semantic retrieval
and LLM orchestration, a custom transformers-based inference server for Param-1, a standalone
Indic Parler-TTS server with streaming chunk synthesis and audio post-processing, and a React +
Vite frontend with real-time token streaming and sequential audio playback — all containerized via
Docker Compose.
What It Demonstrates
Vaani Sahayak illustrates how to:
- Answer natural language queries in Hindi and Telugu using retrieval-augmented generation
over 2,086 pre-embedded government scheme records.
- Serve a custom-architecture LLM (Param-1-2.9B-Instruct) via a transformers-backed
OpenAI-compatible API, bypassing vLLM's architecture restrictions.
- Stream LLM responses token-by-token and stream TTS audio chunk-by-chunk for a low-
latency perceived experience.
- Synthesize Hindi speech from generated text with chunk-based audio production, phantom-
speech removal, and silence trimming.
- Normalize Hindi text for TTS — converting currency symbols, numerals, and English
acronyms into natural spoken Hindi.
- Deploy a full-stack civic AI assistant with optional enterprise GPU inference (Keycloak
OAuth2 + APISIX gateway) or fully local fallback.
- Designed for civic technologists, government innovation teams, AI researchers, and
developers building for Bharat, Vaani Sahayak serves as a reference implementation for
sovereign, multilingual, offline-capable AI assistants for public service delivery.
Key Capabilities
Sovereign AI Stack
Powered entirely by Indian-origin models: Param-1-2.9B-Instruct (BharatGen consortium — IIT Madras, IIT Bombay, IIT Kanpur) for language generation and Indic Parler-TTS (AI4Bharat) for speech synthesis. No OpenAI, no Bhashini, no external cloud dependency.
Hindi and Telugu Support
Citizens query in Hindi or Telugu via a bilingual text interface. Language-specific system prompts and query examples are included; Indic Parler-TTS natively supports both languages for spoken audio responses.
RAG over 2,086 Government Schemes
Cosine similarity search over pre-computed sentence-transformer embeddings of all 2,086 schemes, retrieving top-10 candidates and injecting the top-3 into Param-1's context window for grounded, factual answers.
Real-Time Streaming
Pipeline
LLM tokens stream to the browser via Server-Sent Events as they are generated. TTS audio is synthesized in chunks and played back sequentially as each chunk arrives — delivering a responsive experience despite on-device inference.
Advanced Audio Quality Controls
The TTS server implements chunk-based synthesis, RMS-based silence trimming, phantom- speech detection, and adaptive token budgeting — producing clean, natural-sounding spoken responses.
Flexible Deployment
Runs fully offline without external API dependencies. Optionally integrates with an enterprise GPU inference stack via Keycloak OAuth2 and APISIX gateway for production-grade TTS throughput. Containerized with Docker Compose for reproducible deployment.