🔴 Live Demo:
Sarkari Yojana Sahayak is a voice-first, agentic AI system designed to assist rural users in discovering and applying for government welfare schemes in their native language (Hindi). Unlike simple chatbots, this system uses a Finite State Machine (FSM) via LangGraph to guide the user through a structured conversation flow—from eligibility checks to final application—while handling errors, interruptions, and memory context.
- 🎙️ Voice-First Interface: Complete hands-free experience. The user speaks in Hindi, and the agent replies with a natural, human-like male voice (using Microsoft Edge Neural TTS).
- 🧠 Agentic Workflow: Powered by LangGraph, the system moves through distinct logical stages (Greeting → Data Collection → Scheme Presentation → Application).
- 💾 Contextual Memory: The agent remembers critical user details (Age, Income) across multiple conversation turns.
- 🛡️ Robust Failure Handling: Automatically detects silence (e.g., phantom "Kar do" inputs from Whisper), handles nonsense queries, and politely guides the user back to the correct flow.
- 🔗 Deep Linking: Provides direct, clickable application links to official government portals upon user request.
- Frontend: Streamlit, Streamlit-Mic-Recorder
- Orchestration: LangGraph (State Machine & Graph Theory)
- LLM: Llama-3.3-70b-Versatile (via Groq API)
- Speech-to-Text (STT): Whisper-Large-v3 (via Groq)
- Text-to-Speech (TTS): Microsoft Edge-TTS (
hi-IN-MadhurNeural) - Data Source: JSON-based local scheme database (
schemes.json)
The Sarkari Yojana Sahayak is not a simple chatbot. It is a state-aware agent designed on the Planner-Evaluator-Executor architecture. It uses a Finite State Machine (FSM) implemented via LangGraph to ensure deterministic reliability while maintaining the flexibility of an LLM.
This diagram illustrates the core cognitive loop of the agent. Unlike a standard RAG pipeline, this system evaluates its own outputs before speaking to the user.
- Planner (Perception): The agent listens to the audio, transcribes it, and identifies the user's Intent (e.g., "Giving Income Info") and extracts Entities (Age: 25, Income: 20k).
- Executor (Action): Based on the plan, it executes tools—specifically querying the
schemes.jsondatabase or updating the session memory. - Evaluator (Logic Check): Before responding, the agent checks:
- Did I get the necessary data? (If not, ask follow-up).
- Is the data valid? (If age is 200, reject it).
- Did the tool return results? (If yes, present schemes).
This diagram visualizes the internal code structure within state.py. It shows how the Analyzer Node (LLM) hands off control to the Decision Node (Python Logic) to ensure accurate scheme retrieval.
- Analyzer Node: Uses
Llama-3-70bto understand context and nuance (e.g., correcting "PM Yashwag" to "PM Yashasvi"). - Decision Node: A rigid rule engine that enforces business logic. It prevents the LLM from hallucinating fake schemes.
- State Memory: Persists user data (Age/Income) across multiple conversation turns so the user doesn't have to repeat themselves.
The agent follows a strict Depth-Layered Conversation Flow. This ensures users are guided step-by-step from greeting to application without getting lost.
- Layer 0 (Start): User must initiate with keywords like "Shuru" or "Start".
- Layer 1 (Data Collection): The agent insists on getting valid Age and Income before proceeding.
- Layer 2 (Scheme Presentation): Eligible schemes are listed point-wise.
- Layer 3 (Detail & Application): Deep linking to government portals (
.gov.in) upon user request. - Error Handling: If the user stays silent or speaks nonsense, the agent attempts to realign the conversation twice before politely ending it.
User: "नमस्ते" (Namaste)
Agent: "नमस्ते, आशा करता हूँ आपका दिन अच्छा जा रहा है |" (Status: ✅ Greeting Intent Detected)
User: "मुझे आईस्क्रिम खाना है" (I want to eat ice cream)
Agent: "क्षमा करें, मैं केवल सरकारी योजनाओं में आपकी सहायता कर सकता हूँ।" (Status: 🛡️ Irrelevant Input Rejected)
User: "मेरी उम्र 120 साल है..." (My age is 120...)
Agent: "मनुष्य का औसत जीवनकाल 90 साल होता है, कृपया मुझे अपनी सही उम्र बताएं" (Status: 🛡️ Logic Guardrail - Biological Impossibility Rejected)
User: "अच्छा, मेरी उम्र 15 साल है" (Okay, my age is 15)
Agent: "सही योजना खोजने के लिए मुझे आपकी उम्र और आय दोनों की आवश्यकता होगी |" (Status: 🧠 Memory - Retained Age, Planner requested missing Income)
User: "मेरी आय 5000 है" (My income is 5000)
Agent: "आपकी जानकारी के आधार पर, आप निम्नलिखित योजनाओं के लिए पात्र हैं:-
1. अल्पसंख्यक प्री-मैट्रिक छात्रवृत्ति
2. पीएम यशस्वी छात्रवृत्ति योजना..." (Status: 🛠️ Tool Call - Database Search Executed)
User: "पियन यशाश्री चात्रपृत्त योजना" (Mispronounced Scheme Name)
Agent: "पीएम यशस्वी छात्रवृत्ति योजना: OBC, EBC और DNT श्रेणी के मेधावी छात्रों के लिए उच्च स्तरीय शिक्षा छात्रवृत्ति। क्या आप आवेदन करना चाहते हैं?" (Status: 🤖 AI Reasoning - Fuzzy Matching Corrected 'Pian Yashashri' to 'PM Yashasvi')
User: (Silence/Background Noise detected as text"कर दो" )
Agent: "माफ़ करें, मैं आपको समझ नहीं पाया। क्या आप कृपया फिर से दोहरा सकते हैं?" (Status: 🔁 Failure Handling - Null/Ambiguous Input Detected)
User: "जी हाँ" (Yes)
Agent: "बढ़िया! आप इस लिंक पर जाकर आवेदन कर सकते हैं: [यहाँ क्लिक करें]" (Status: 🛠️ Tool Call - Application Link Retrieved)
git clone https://github.com/BleeGleeWee/Voice-Based-Native-Language-Service-Agent-
cd Voice-Based-Native-Language-Service-Agent-
pip install -r requirements.txt
python -m venv agent_env
agent_env\Scripts\activate
Create a file named .streamlit/secrets.toml in the root directory.
⚠️ Security Warning: Never upload this file to GitHub. Add it to your.gitignore.
GROQ_API_KEY = "your_groq_api_key_here"
streamlit run streamlit_app.py
Sarkari-Yojana-Sahayak/
├── .streamlit/
│ └── secrets.toml # API keys configuration (NOT pushed to GitHub)
├── schemes.json # Knowledge Base (Eligibility Rules & Data)
├── state.py # Core Agent Logic (LangGraph State Machine)
├── streamlit_app.py # Frontend Interface (UI, Audio I/O)
├── stt.py # Speech-to-Text Utility
├── requirements.txt # Project Dependencies
└── README.md # Documentation