Bharat Benefits Voice Agent

A voice-first RAG assistant for Indian public-benefit schemes, built with Moss semantic retrieval and Sarvam AI speech APIs.

---

What This Does

Users ask questions by voice or text about Indian government welfare schemes and receive spoken, grounded answers in return.

voice / text input
    |
Sarvam STT  (saaras:v3)         converts speech to text
    |
Moss retrieval                  finds relevant scheme snippets from local knowledge base
    |
Sarvam Chat Completion          generates a grounded, voice-friendly answer
    |
Sarvam TTS  (bulbul:v3)         converts the answer to speech
    |
audio response                  played back through speakers

The knowledge base covers five schemes:

• Ayushman Bharat / PM-JAY - health insurance for poor families
• PM-KISAN - income support for small farmers
• PM Vishwakarma - skill and financial support for artisans
• National Scholarship Portal - centralised scholarship access for students
• PMKVY / Skill India - free vocational skill training and certification

This is a demo knowledge base, not an official government service. Always verify eligibility, amounts, and deadlines on the official government portals listed in each scheme file.

---

Recording Modes

The agent supports three input modes:

Fixed window (default) Records for a set number of seconds per turn. Simple and predictable.

python bot.py              # 5 seconds per turn
python bot.py --seconds 8  # 8 seconds per turn

VAD mode (recommended for natural conversation) Uses voice activity detection to automatically stop recording when you pause speaking. No need to worry about a timer — just speak naturally and the agent picks up when you stop.

python bot.py --vad

Text mode (for debugging) Type questions instead of speaking. Useful when testing without a microphone.

python bot.py --text-mode

---

Why Moss?

Moss is a real-time semantic search runtime for conversational AI agents.

• Sub-10 ms retrieval, fast enough for a live voice turn.
• Simple SDK: create_index then load_index then query. No cluster to manage.
• Hybrid search (semantic + keyword) with a single alpha parameter.
• Cloud sync: index once, load anywhere.

Moss retrieves the top-3 most relevant scheme snippets for each question, which are passed as grounded context to the LLM. This prevents hallucination of eligibility rules, amounts, or deadlines.

Moss SDK methods used

Method	Purpose
MossClient(project_id, project_key)	Initialise the client
client.create_index(name, documents)	Index scheme files (run once)
client.load_index(name)	Load index into local runtime before querying
client.query(name, question, QueryOptions(top_k=3, alpha=0.75))	Retrieve top-3 relevant snippets

alpha=0.75 blends 75% semantic search with 25% keyword search.

---

Why Sarvam?

Sarvam AI is India's purpose-built AI stack for Indian languages.

• Saaras v3 (STT): speech-to-text with support for 23 Indian languages including Hindi, Bengali, Tamil, Telugu, Punjabi, and more.
• Sarvam-30B / 105B (Chat): multilingual chat models with Indian linguistic context.
• Bulbul v2/v3 (TTS): natural-sounding Indian-language voices.

The combination handles queries in multiple Indian languages out of the box.

---

Setup

1. Navigate to the demo folder

cd moss-live-labs/community-demos/voice-agents/bharat-benefits/

2. Create a virtual environment

python -m venv .venv
source .venv/bin/activate          # Linux / macOS
# .venv\Scripts\activate           # Windows

3. Install dependencies

pip install -r requirements.txt

On Debian/Ubuntu, if simpleaudio fails to build:

sudo apt-get install libasound2-dev

webrtcvad is only required for --vad mode. If you don't need it, the rest of the agent works without it.

4. Set environment variables

cp .env.example .env

Fill in:

Variable	Where to get it
SARVAM_API_KEY	dashboard.sarvam.ai
MOSS_PROJECT_ID	moss.dev
MOSS_PROJECT_KEY	moss.dev

The other variables have sensible defaults.

5. Create the Moss index (run once)

python create_index.py

Expected output:

Loading scheme documents from: data/schemes/
Found 5 file(s):
  ayushman-bharat.md
  national-scholarship-portal.md
  pm-kisan.md
  pm-vishwakarma.md
  pmkvy-skill-india.md

Creating Moss index 'bharat-benefits' ...
Index 'bharat-benefits' created with 5 document(s).

---

Running the Agent

# Fixed window, 5 seconds per turn (default)
python bot.py

# Fixed window, custom duration
python bot.py --seconds 8

# VAD mode - stops when you pause (most natural)
python bot.py --vad

# Text mode - type questions (no mic needed)
python bot.py --text-mode

---

Example Questions

Question	Relevant scheme
"I am a small farmer. Which scheme can help me?"	PM-KISAN
"What documents are needed for Ayushman Bharat?"	Ayushman Bharat
"I am a carpenter. Is there any government scheme for me?"	PM Vishwakarma
"How much money does PM-KISAN give, and how often?"	PM-KISAN
"I want free vocational training. What should I do?"	PMKVY
"My family is poor and needs medical coverage."	Ayushman Bharat
"I am a student looking for scholarship support."	NSP
"Which scheme is useful for skill training?"	PMKVY / Skill India

---

Environment Variables

SARVAM_API_KEY          Sarvam AI subscription key (required)
MOSS_PROJECT_ID         Moss project ID (required)
MOSS_PROJECT_KEY        Moss project key (required)
MOSS_INDEX_NAME         Index name, default: bharat-benefits (optional)
SARVAM_STT_MODEL        STT model, default: saaras:v3 (optional)
SARVAM_CHAT_MODEL       Chat model, default: sarvam-105b (optional)
SARVAM_TTS_MODEL        TTS model, default: bulbul:v3 (optional)
SARVAM_TTS_SPEAKER      TTS speaker voice, default: priya (optional)

---

Folder Structure

bharat-benefits/
├── README.md
├── .env.example
├── requirements.txt
├── create_index.py        one-time Moss index creation
├── bot.py          main voice agent (STT -> Moss -> LLM -> TTS)
└── data/
    └── schemes/
        ├── ayushman-bharat.md
        ├── pm-kisan.md
        ├── pm-vishwakarma.md
        ├── national-scholarship-portal.md
        └── pmkvy-skill-india.md

---

Limitations

• Demo knowledge base only. Scheme information may not reflect the latest government guidelines, benefit amounts, or eligibility criteria.
• Not an official government service. Cannot process applications, verify beneficiary status, or access real-time government databases.
• Audio playback requires simpleaudio. If unavailable, audio is saved to disk.
• VAD mode requires webrtcvad. Falls back to fixed-window if not installed.

Official portals for verification:

• Ayushman Bharat: https://pmjay.gov.in
• PM-KISAN: https://pmkisan.gov.in
• PM Vishwakarma: https://pmvishwakarma.gov.in
• National Scholarship Portal: https://scholarships.gov.in
• Skill India / PMKVY: https://www.skillindia.gov.in

---

Dependencies

Package	Version	Purpose
moss	>=1.0.0	Moss semantic search SDK
python-dotenv	>=1.0.0	Load .env credentials
httpx	>=0.27.0	HTTP client for Sarvam API calls
sounddevice	>=0.4.6	Microphone recording and playback
soundfile	>=0.12.1	WAV file read/write
simpleaudio	>=1.0.4	Audio playback (optional)
numpy	>=1.26.0	Audio array handling
webrtcvad	>=2.0.10	Voice activity detection for --vad mode (optional)

---

Built for the Moss Live Labs community.