nano-graphrag

Why we included this project

Microsoft's official GraphRAG pipeline is powerful, but tracing how it actually works can feel like wandering through a maze. nano-graphrag compresses that approach into roughly a thousand lines of readable, fully typed Python while keeping the core graph-based retrieval flow: it extracts entities and relations from your documents, builds a knowledge graph, clusters it into communities, and generates reports you can query in global or local mode. That makes it a solid reference for teams that want to understand graph retrieval before committing to a heavier production stack. It is also practical to run: OpenAI, Bedrock, and local models via Ollama all work as backends, and for vector storage you can choose the built-in option, faiss, hnswlib, or Milvus, with Neo4j available for graph storage. Developers building a RAG feature who want something they can read end-to-end and adapt to their own data will find this far easier to follow than the official implementation.