Teams doing RAG over large document sets who need better global reasoning than plain vector search.

A modular graph-based RAG system from Microsoft Research that extracts entity and relationship knowledge graphs from text corpora to improve retrieval quality and holistic question answering.

Runs as a Python pipeline that builds a graph index from raw documents, queryable afterward via LLM calls.

Use when questions require synthesizing information across many documents rather than retrieving a single relevant chunk.

It answers broad, thematic questions that standard chunk-based vector retrieval tends to miss.

Depends on an LLM for entity extraction and summarization plus a graph storage backend.

Open source, MIT licensed, published and maintained by Microsoft Research; free to run with your own LLM key.