Methods

Method

GraphRAG

Retrieval method that builds an entity-relation graph from the corpus, summarises graph communities, and uses those summaries plus graph traversal alongside vector search.

Source

https://arxiv.org/abs/2404.16130

Taxonomy

methods.rag

Origin

Microsoft Research (2024)

Primary source

https://arxiv.org/abs/2404.16130

Domain

rag

Maturity

established

Primary artifacts

entity graph, community summaries, hierarchical clusters

Core idea

Naive RAG retrieves chunks by vector similarity. GraphRAG adds a structural layer:

1. Extract entities and relations from each chunk with an LLM pass.
2. Cluster the graph into communities (Leiden/Louvain).
3. Summarise each community with an LLM — these summaries become first-class retrievable units alongside raw chunks.
4. Retrieve by combining vector hits, graph neighbourhood traversal, and matching community summaries.

Why it matters

Vector search alone struggles with global questions (“what are the main themes across this corpus?”) because the answer is not in any single chunk. Community summaries give the retriever a coarse-to-fine ladder.

Implementations

• Microsoft GraphRAG — reference implementation, heaviest.
• LightRAG — simpler, fewer LLM calls, pluggable backends.
• Neo4j GraphRAG — uses Neo4j as the graph store.

Trade-offs

• Ingest cost is meaningfully higher than naive RAG (one LLM pass per chunk for extraction plus one per community for summarisation).
• Quality gain is largest on corpora with strong relational structure and on global questions; on flat FAQ-style content the lift is marginal.