PolicyMind

A Graph-RAG chatbot that lets you upload internal policy documents and ask natural language questions about them — such as "What do I need to consider before launching a new vendor relationship?"

Built with LangChain, Neo4j, Ollama, and FastAPI. React frontend coming soon.

---

Why Graph-RAG?

Standard RAG retrieves isolated text chunks. PolicyMind uses a knowledge graph to capture relationships between policies, topics, and dependencies. This means answers are more complete — if Policy A references Policy B, the system understands that connection.

---

Features

• Upload PDF or Markdown policy documents via API
• Automatic extraction of entities and relationships into Neo4j
• Natural language Q&A grounded in your documents
• Source citations for every answer
• Runs fully locally — no API key required
• Dynamic: swap in any document set without code changes

---

How it works

1. Documents are ingested and split into chunks.
2. Each chunk is embedded using Ollama (nomic-embed-text) and stored as a node in Neo4j.
3. Chunks are linked sequentially via NEXT relationships and grouped under a Document node.
4. At query time, a vector similarity search retrieves the most relevant chunks.
5. The context is passed to a local LLM via Ollama (qwen3:8b), orchestrated by LangChain.
6. The model generates a grounded answer with references to the source documents.

---

Architecture

flowchart TB
    subgraph Ingestion["Ingestion Pipeline"]
        PDF["PDF / Text File"]
        SPLIT["Text Splitter\nRecursiveCharacterTextSplitter"]
        EMBED_I["Ollama Embeddings\nnomic-embed-text"]
        PDF --> SPLIT --> EMBED_I
    end

    subgraph Neo4j["Neo4j Knowledge Graph"]
        DOC["Document Node"]
        CHUNK1["Chunk Node 0"]
        CHUNK2["Chunk Node 1"]
        CHUNK3["Chunk Node n"]
        DOC -->|HAS_CHUNK| CHUNK1
        DOC -->|HAS_CHUNK| CHUNK2
        DOC -->|HAS_CHUNK| CHUNK3
        CHUNK1 -->|NEXT| CHUNK2
        CHUNK2 -->|NEXT| CHUNK3
    end

    EMBED_I -->|store embedding + text| Neo4j

    subgraph Query["Query Pipeline"]
        USER["User Question"]
        EMBED_Q["Ollama Embeddings\nnomic-embed-text"]
        RETRIEVER["Vector Similarity Search\nTop-k Chunks"]
        LLM["Ollama LLM\nqwen3:8b"]
        ANSWER["Answer + Sources"]
        USER --> EMBED_Q --> RETRIEVER --> LLM --> ANSWER
    end

    Neo4j -->|retrieve relevant chunks| RETRIEVER

    subgraph API["API Layer"]
        FASTAPI["FastAPI\nPOST /ask\nPOST /upload"]
    end

    USER -.->|HTTP Request| FASTAPI
    FASTAPI -.-> USER
    FASTAPI --> Query

Loading

Tech Stack

Layer	Technology
Orchestration	LangChain
Graph Database	Neo4j
Embeddings	Ollama (nomic-embed-text)
LLM	Ollama (qwen3:8b)
Backend API	FastAPI
Frontend	React (in progress)

---

Project Structure

policymind/
├── backend/
│   ├── api/          # FastAPI routes
│   ├── core/         # LangChain chains and RAG logic
│   ├── graph/        # Neo4j ingestion and query logic
│   └── models/       # Pydantic schemas
├── docs/             # Example policy documents
├── scripts/          # Ingestion and setup scripts
├── docker-compose.yml
└── README.md

---

Getting Started

Prerequisites

• Python 3.11+
• Docker and Docker Compose
• Ollama with nomic-embed-text and qwen3:8b pulled

ollama pull nomic-embed-text
ollama pull qwen3:8b

Setup

git clone https://github.com/your-username/policymind.git
cd policymind

cp .env.example .env
# Edit .env and set your Neo4j password

docker compose up -d
pip install -r requirements.txt

PYTHONPATH=. python scripts/ingest.py --file docs/example_policy.pdf --name "Example Policy"
uvicorn backend.api.main:app --reload

Example Query

curl -X POST http://localhost:8000/ask \
  -H "Content-Type: application/json" \
  -d '{"question": "What do I need to consider before onboarding a new vendor?"}'

{
  "answer": "According to the Vendor Management Policy, you must complete a risk assessment, obtain approval from the procurement team, and ensure GDPR compliance before onboarding a new vendor.",
  "sources": ["vendor_management_policy.pdf", "data_privacy_guidelines.pdf"]
}

---

Roadmap

Core RAG System

• Document ingestion pipeline (PDF / Text → Chunking → Neo4j storage)
• Graph structure with sequential chunk linking (NEXT relationships)
• Vector-based retrieval with Neo4j (embedding similarity search)
• LLM-based question answering via RAG pipeline
• Upgrade to true Graph-RAG (relationship-aware retrieval + multi-hop traversal)
  • Use Neo4j relationships (e.g. NEXT, RELATED, HAS_ENTITY) during retrieval
  • Add graph-based context expansion after vector search
  • Implement hybrid retrieval (vector + graph traversal)
  • Add reranking of expanded context for better answer quality

Backend API

• FastAPI Q&A endpoint (production-ready)
• Structured response format (answer + sources + context)
• Streaming responses for real-time output

Frontend

• React frontend with document upload UI
• Chat interface for querying policies
• Source highlighting and traceable answers

System Features

• Multi-tenant support
• Authentication and role-based access control
• Document versioning and updates

Infrastructure

• Dockerized full-stack setup
• CI/CD pipeline
• Observability (logging and tracing for RAG pipeline)

---

License

MIT