GraphRAG — Knowledge Graph-Based Retrieval-Augmented Generation

Definition

GraphRAG is a next-generation RAG pattern that converts documents into a knowledge graph and then uses the graph structure to retrieve information.

TL;DR

GraphRAG is a RAG extension pattern released by Microsoft in 2024 that structures documents as an entity-relationship graph rather than as simple chunks. Whereas basic RAG relies on the semantic similarity of individual document chunks, GraphRAG understands the relationships between entities to answer complex questions accurately.

Limitations of Basic RAG

Basic RAG retrieves text chunks that are semantically similar to the question. It is effective for simple questions ("What is AEO?") but weak on complex questions like these:

• "What other companies did Company A's CEO found?" (multi-step relationships)
• "What are the overall trends in this field?" (global summarization)
• "How does the relationship between X and Y affect Z?" (relational reasoning)

Answering such questions requires understanding the relationships between pieces of information rather than individual chunks.

How GraphRAG Works

GraphRAG works in three stages.

Stage 1: Graph construction
Extract entities (people, organizations, concepts) from documents and structure the relationships between entities as a graph. "Google launched AI Overviews" → entities: Google, AI Overviews / relationship: launched.

Stage 2: Community detection
Identify groups of closely connected entities (communities) within the graph. Generate a summary for each community to build a hierarchical index.

Stage 3: Question answering
When a question comes in, find the relevant communities and entities and follow the graph structure to generate an answer.

Implications from an AEO Perspective

For content to be well utilized in a GraphRAG system:

• Define entities clearly: "X is a sub-concept of Y," "Tool A is used for task B"
• State relationships explicitly: Write comparison tables, taxonomies, and causal relationships in a structured way
• Apply structured data (JSON-LD): Organization, Person, and DefinedTerm schemas help with entity recognition

Frequently Asked Questions

Q. Do ordinary site owners need to build GraphRAG themselves?
A. No. GraphRAG is an internal structure of the AI system. Content creators only need to describe entities and relationships clearly so that a GraphRAG system can process the content well.

Related Sources

• Microsoft Research (2024). GraphRAG: Unlocking LLM discovery on narrative private data. https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/
• Edge et al. (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. https://arxiv.org/abs/2404.16130