This project provides a lightweight and hackable implementation of GraphRAG, designed for developers and researchers seeking a simpler alternative to existing complex frameworks. It offers core GraphRAG functionality with a focus on modularity, portability, and ease of customization, enabling faster iteration and integration into various RAG applications.

How It Works

nano-graphrag constructs a knowledge graph from text, enabling more sophisticated retrieval than traditional RAG. It extracts entities and relationships to build this graph, allowing for both "global" (whole-graph) and "local" (context-specific) searches. This approach aims to improve relevance and context understanding by leveraging structured knowledge, with a design prioritizing small code size (around 1100 lines excluding tests/prompts) and asynchronous operations.

Quick Start & Requirements

• Install: pip install -e . (from source, recommended) or pip install nano-graphrag (from PyPI).
• Prerequisites: OpenAI API key (or configure Azure OpenAI/Amazon Bedrock). Examples are provided for using Ollama or transformers.
• Setup: Requires setting API keys and potentially downloading data.
• Docs: README, Advances, FQA, Roadmap.

Highlighted Details

• Supports multiple LLM providers (OpenAI, Azure OpenAI, Amazon Bedrock, DeepSeek, Ollama) and embedding models (OpenAI, Sentence-transformers).
• Integrates with various vector databases (nano-vectordb, hnswlib, milvus-lite, faiss) and graph storage (networkx, Neo4j).
• Offers both global and local search modes, with customizable chunking and prompt management.
• Asynchronous API support for all core methods.

Maintenance & Community

• Actively maintained with contributions welcomed.
• Benchmarks and evaluation notebooks are available.
• Projects using nano-graphrag are listed, encouraging community adoption.

Licensing & Compatibility

• The specific license is not explicitly stated in the README, but the project structure and examples suggest a permissive open-source license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The project does not implement the "covariates" feature found in the original GraphRAG. Its global search implementation differs by using only the top-K important and central communities, rather than a map-reduce style across all communities.