Agentic GraphRAG System

An advanced, agentic Hybrid GraphRAG system designed to extract and retrieve information from documents using a combination of vector search and knowledge graphs. Built for high accuracy and deep, multi-hop reasoning over complex text.

Architecture Overview

This project implements a dual-retrieval pipeline managed by a stateful agentic loop:

1. Vector Retrieval: Uses ChromaDB to store and retrieve semantic chunks. Embeddings are generated locally using the Alibaba-NLP/gte-Qwen2-1.5B-instruct model via sentence-transformers.
2. Graph Retrieval: Uses an LLM to extract entities and relationships, storing them in a local NetworkX graph. This enables multi-hop reasoning across connected concepts.
3. Agentic Loop: Utilizes LangGraph (specifically a ReAct agent) to dynamically route queries between the Vector Database and the Knowledge Graph. If the initial retrieved context is insufficient, the agent self-reflects, reformulates its query, and loops until a satisfactory answer is formed.

Prerequisites

• Python 3.9 or higher
• 32GB RAM recommended for CPU-only environments.

Installation

1. Clone the repository and navigate to the project directory:

git clone <your-repo-url>
cd DocRag

2. Create a virtual environment and activate it:

python -m venv venv
# On Windows:
.\venv\Scripts\activate
# On Linux/Mac:
source venv/bin/activate

3. Install the required dependencies:

pip install -r requirements.txt

Configuration

By default, the system is configured to use local CPU inference.

Local LLM via Ollama

If running purely locally, ensure you have Ollama installed and the Qwen 2.5 model pulled:

ollama pull qwen2.5

Note: Extracting the knowledge graph purely on a CPU via a local LLM can be extremely time-consuming.

High-Speed API (Recommended for Graph Indexing)

To significantly accelerate the knowledge graph extraction phase, you can use a free-tier API such as Together AI. Create a .env file in the root directory and add your API key:

TOGETHER_API_KEY=your_api_key_here

When this key is present, the system will automatically switch to the remote API for LLM inference while keeping vector embeddings strictly local.

Usage

1. Ingesting a Document

To parse a PDF, chunk the text, generate embeddings, and build the knowledge graph, run the ingestion script:

python ingest.py "path/to/your/document.pdf"

This process will create a chroma_db directory for the vectors and a graph_db.graphml file for the knowledge graph.

2. Querying the System

Once the document is ingested, start the agentic chat interface:

python main.py

Type your questions into the prompt. The LangGraph agent will output its reasoning process as it searches the vector database and the knowledge graph to synthesize the final answer. Type exit or quit to terminate the session.

License

This project is licensed under the MIT License.