Skip to content
BENIBIX Learn

GraphRAG

Also known as Graph RAG, Graph Retrieval-Augmented Generation, Knowledge Graph RAG

RAG that makes a map of facts from your files. Then it searches that map too.

Think of it like

It is like a map of who knows who in a town. Big questions get easy.

Example

Microsoft Research put its GraphRAG on GitHub in July 2024. A team uses it to ask "what are the big risks across these 500 reports?"

Why it matters

Some questions span a whole collection, not one passage. A graph with summaries can answer them.

Definition

GraphRAG is RAG that searches a knowledge graph as well as text chunks. First, a model reads every document and pulls out the people, places and links. These form a graph. The graph is split into groups of linked things, and each group gets a summary. Questions about the whole collection can then use those summaries. Plain RAG struggles with "what are the main themes here?"
GraphRAG builds this map from the text, then sums up each cluster.

Common mistake

Using it for simple fact lookups. Building the graph costs many model calls, and plain RAG already does that well.

Key takeaways

  1. 1
    Build a graph. Things and links from the text.
  2. 2
    Summarize groups. Each cluster gets an overview.
  3. 3
    Big questions. Themes across many documents.