GraphRAG (Knowledge-Graph Retrieval)

What is it?

Ordinary RAG fetches a few similar passages. Some questions need connections — "how are these people, projects and events related?" — or an overview of a whole collection. GraphRAG builds a map of entities and links first, then uses it to answer such questions.

Also known as: GraphRAG, graph RAG, knowledge graph RAG, knowledge graphs for LLMs, multi-hop retrieval, community summaries, Neo4j RAG, entity relationship retrieval.

Technical definition

Microsoft Research's "From Local to Global" (Edge et al., 2024) describes building an entity graph from text with an LLM, detecting communities, pre-summarising them, and answering global questions via map-reduce over summaries; "GraphRAG" also denotes broader graph-based retrieval approaches (graph databases, multi-hop traversal). Costs include LLM extraction at index time, graph quality issues (entity resolution) and update complexity. Gains are task-specific and must be validated.

Entity type: technique · Freshness class: volatile.

When should I use it?

Use for sensemaking over large corpora and relationship-heavy questions.

Prerequisites and related concepts

Understand these first: What is RAG (Retrieval-Augmented Generation)?, What are Embeddings in AI?.

Related concepts: Graph Neural Networks (GNNs), Hybrid Search and Reranking, Agentic RAG (Adaptive and Self-Correcting Retrieval), What is a Vector Database?, RAG Evaluation (Retrieval and Answer Quality).

Implementation patterns

• Extract entities/relations with an LLM; resolve duplicates.
• Community detection and summarisation for global questions.
• Hybrid: vector retrieval for local questions, graph for global.
• Cache or incrementally update the graph.

Technologies and frameworks

Commonly associated: Microsoft GraphRAG, Neo4j, LlamaIndex property graphs, entity-resolution tooling. Names are examples of what exists, not recommendations, and projects change quickly.

Practical guide

1. Check if questions need synthesis or multi-hop reasoning.
2. Pilot on a subset.
3. Audit extracted graph quality.
4. Compare against hybrid RAG on your questions.

Common failure modes

• Extraction errors creating false relationships.
• High indexing cost.
• Stale graph after document updates.
• Using it where plain RAG would do.

Troubleshooting

• Wrong or duplicate entities — Improve entity resolution and prompts; audit samples.
• Cost too high — Limit to the document subsets that need graph answers.

Contrasted with

• What is RAG (Retrieval-Augmented Generation)? — Standard RAG retrieves similar chunks; GraphRAG retrieves connected structure and summaries.
• Graph Neural Networks (GNNs) — GNNs are learned models over graphs; GraphRAG uses graph structure to organise LLM retrieval.

Questions this page answers

• What is GraphRAG?
• GraphRAG vs vector RAG?
• When do I need a knowledge graph for RAG?
• How do I build a knowledge graph from documents?

Typical goals: answer multi-hop or corpus-wide questions; decide whether GraphRAG is worth the cost.

Sources and verification

Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.

Freshness class: volatile — Products, versions, protocols, benchmarks, leaderboards, regulations and dates. Verify against the primary source before relying on specifics; re-review every 3 months.

Time-sensitive claims to verify before relying on them:

• Tooling names, repository status and cost characteristics.

Canonical sources (unverified leads):

• Edge et al., From Local to Global: A Graph RAG Approach to Query-Focused Summarization (2024) — arXiv:2404.16130

Related concepts