Graph Policy - AI Powered Policy Assistant

Graph Policy is an enterprise-grade AI assistant capable of processing complex government schemes, modeling them dynamically using Knowledge Graphs, and conducting natural language Retrieval Augmented Generation (RAG) search combined with deterministic Eligibility Evaluation.

Complete Architecture Overview

This is a production-ready rewrite supporting completely dynamic, data-driven reasoning.

1. Ingestion Pipeline (ingestion/pipeline.py)
   • Downloads, parses, and chunks raw PDF, HTML, and TXT data natively using LangChain loaders.
   • Leverages Heuristics or Google Gemini (if configured in .env) to parse explicit Scheme Nodes, Benefits, Criterions, and State structures into our Knowledge Graph without hardcoding anything.
2. Retrieval Engine (retrieval/vector_store.py)
   • Implements native FAISS Vector indexing using HuggingFaceEmbeddings (all-MiniLM-L6-v2) to capture context semantics and exact document sources.
3. Dynamic Knowledge Graph (kg/graph_store.py)
   • Maintains an InMemoryGraphStore via NetworkX mapping relationships like HAS_ELIGIBILITY and PROVIDES_BENEFIT.
   • Serializes directly to data/indices/graph.json so memory states persist.
4. Data-Driven Eligibility (eligibility/engine.py)
   • Completely removes hardcoded evaluations.
   • Evaluates a User Profile dynamically by fetching raw Criterion Nodes directly from the Knowledge Graph on a per-scheme basis and parsing relational operators (e.g. <=, in).
5. Modern API Backend (api/server.py)
   • Offers robust endpoints like /query (runs combined graph traversal + FAISS retrieval + LLM synthesis) and /eligibility (dynamic rule checking).
6. Next-Generation React UI
   • Built with Vite, Tailwind CSS v4, Framer Motion, and Recharts.
   • Includes real dashboard analytics, knowledge graph visualizations via react-force-graph, full Perplexity-like Search, and multi-colored Eligibility trace outputs.

---

🚀 Getting Started

1. Backend Setup & Ingestion

# Install dependencies
pip install -r requirements.txt pydantic-settings

# Start the Ingestion Pipeline. If no data exists, this will auto-seed
# the data/raw/ directory with PM-KISAN, Ayushman Bharat, PMAY, Startup India etc.,
# completely process them, build FAISS embeddings, and construct the KG.
python ingestion/run_ingest.py

# Optional: Ensure Gemini API is available for superior extractions and synthesize
# Add variable back into .env file if necessary.
echo "GEMINI_API_KEY=your_key" > .env

# Run the FastAPI server
uvicorn api.server:app --reload

API will run on http://localhost:8000

2. Frontend Setup

cd ui/frontend
npm run dev

Vite Server will run cleanly on http://localhost:5173

3. Evaluation

# Measure precision, recall, and F1 on document retrieval 
python evaluation/eval.py

4. Deployment Checklists

Option A: Docker Compose (All-in-one)

The repo contains a Dockerfile.

1. docker build -t graph-policy-backend .
2. docker run -p 8000:8000 graph-policy-backend

Option B: Render / Railway / AWS (Backend)

• Standard Uvicorn initialization.
• Build Command: pip install -r requirements.txt
• Start Command: uvicorn api.server:app --host 0.0.0.0 --port $PORT

Option C: Vercel / Netlify (Frontend)

• Build Command: cd ui/frontend && npm install && npm run build
• Output Directory: dist