NagrikAI

AI-powered Indian Business Compliance & Advisory Agent.

Vision

Evolve from a GST RAG chatbot into a Business Compliance & Advisory Agent that diagnoses business problems, gathers missing facts, reasons over them, retrieves evidence, and produces actionable plans.

RAG answers questions. An agent diagnoses a business problem, gathers missing facts, reasons over them, retrieves evidence, and produces an actionable plan.

Why NagrikAI

                    Business Owner
                         │
                         ▼
                ┌─────────────────┐
                │ Business Agent  │
                │   Orchestrator  │
                └────────┬────────┘
                         │
          ┌──────────────┼──────────────┐
          ▼              ▼              ▼
     Problem         Business        Knowledge
     Detection       Context         Retrieval
          │              │              │
          ▼              ▼              ▼
   ┌──────────┐   ┌─────────────┐ ┌──────────────┐
   │ GST RAG  │   │ User/Business│ │ Gov / Legal  │
   │          │   │ Profile      │ │ Documents    │
   └──────────┘   └─────────────┘ └──────────────┘
          │              │              │
          └──────────────┼──────────────┘
                         ▼
                  ┌─────────────┐
                  │ Reasoning / │
                  │ Planning    │
                  └──────┬──────┘
                         ▼
                  ┌─────────────┐
                  │ Action Plan │
                  └─────────────┘

Problem Domains

1. GST & Taxation

• GST registration, composition scheme, input tax credit, GST returns
• E-invoicing, e-way bills, HSN/SAC, reverse charge
• GST notices, registration amendments, cancellation, refunds

2. Business Registration & Government Compliance

• MCA, Udyam, Income Tax, GST, Shops & Establishments
• EPFO, ESIC, FSSAI, DGFT, state-specific registrations, local licenses

3. Business Decision Support

• Revenue/margin analysis, hiring affordability, EMI calculations
• Tools: calculator, financial_analysis, tax_calculator

4. Government Scheme Discovery

• Subsidies, incentives, loans, and schemes based on industry, location, company size, investment, employment

5. Compliance Calendar

• GST: GSTR-1, GSTR-3B, Annual return
• Income Tax: Advance tax
• MCA: AOC-4, MGT-7
• Labour: PF, ESIC

Bounded context, on purpose

NagrikAI operates with a finite context window determined by the LLM's capacity (typically 8K-32K tokens). Rather than attempting to retrieve and reason over unlimited documents, we:

• Retrieve top-K chunks (configurable, default 5) after reranking from Fetch-K candidates (default 20)
• Apply strict token limits to context construction
• Truncate overly long documents during parsing
• Accept that some niche queries may require refinement or fallback to web search

This constraint ensures predictable latency, prevents context overflow errors, and focuses the LLM on the most relevant information – mirroring how human researchers work with limited cognitive bandwidth.

Installation

# Clone and install
git clone https://github.com/your-org/nagrik-ai.git
cd nagrik-ai

# Install with UV (recommended)
uv sync

# Or with pip
pip install -e .

# Verify installation
uv run nagrik-ai --help

Quick start

# 1. Crawl official government sources (runs scrapy spiders)
uv run nagrik-ai crawl sites

# 2. Parse HTML to Markdown
uv run nagrik-ai parse all

# 3. Vectorize documents into ChromaDB
uv run nagrik-ai vectorize

# 4. Launch the Gradio web interface
uv run nagrik-ai app-command

# Optional: Use OpenRouter instead of local Ollama
uv run nagrik-ai app-command --llm-provider openrouter

Commands

Command	Description
nagrik-ai crawl sites	Crawl all configured government sites
nagrik-ai parse all	Convert crawled HTML to clean Markdown
nagrik-ai vectorize	Generate embeddings and populate ChromaDB
nagrik-ai app-command	Launch the Gradio web UI
nagrik-ai trace test "Your question"	Test with LangSmith tracing enabled
uv run pytest tests/ -v	Run the test suite

Shell integration

Add this to your shell profile (~/.bashrc, ~/.zshrc, etc.) for quick access:

# Quick NagrikAI access
alias nagrik='uv run nagrik-ai app-command'

Then simply type nagrik in your terminal to launch the assistant.

Supported providers

Provider	Setup	Notes
Ollama (default)	ollama serve + ollama pull mistral:latest	Fully private, local inference
OpenRouter	Set NAGRIKAI_OPENROUTER_API_KEY env var	Access to 100+ cloud LLMs

Example configuration

Create .env file (or set environment variables):

# LLM Provider (ollama or openrouter)
NAGRIKAI_LLM_PROVIDER=ollama

# Ollama settings (used when provider=ollama)
NAGRIKAI_OLLAMA_BASE_URL=http://localhost:11434
NAGRIKAI_OLLAMA_MODEL=mistral:latest

# OpenRouter settings (used when provider=openrouter)
NAGRIKAI_OPENROUTER_API_KEY=your-key-here
NAGRIKAI_OPENROUTER_MODEL=anthropic/claude-3.5-sonnet

# Embedding and reranking
NAGRIKAI_EMBEDDING_MODEL=BAAI/bge-m3
NAGRIKAI_RERANKER_MODEL=BAAI/bge-reranker-large
NAGRIKAI_RERANKER_ENABLED=true
NAGRIKAI_HYBRID_SEARCH_ENABLED=true

# Retrieval tuning
NAGRIKAI_TOP_K=5
NAGRIKAI_FETCH_K=20
NAGRIKAI_LAMBDA_MULT=0.7
NAGRIKAI_BM25_K1=1.5
NAGRIKAI_BM25_B=0.75
NAGRIKAI_RRF_K=60

# Observability
LANGSMITH_TRACING_ENABLED=true
LANGSMITH_API_KEY=your-langsmith-key
LANGSMITH_PROJECT=nagrik-ai

License

MIT

Acknowledgements

• LangChain and LangGraph for LLM orchestration
• ChromaDB for vector storage
• Sentence Transformers for embedding models
• BM25 via rank-bm25 for keyword search
• Gradio for the web interface
• Typer for the CLI
• Pydantic for configuration management
• Ruff and MyPy for code quality
• Official Indian government websites (india.gov.in, gst.gov.in, tutorial.gst.gov.in) as knowledge sources