## Definition **GraphRAG** is a [[Retrieval-Augmented Generation]] variant that builds a knowledge graph (entities and their relationships) from a corpus and retrieves over its **structure**, rather than fetching independent chunks by similarity. Introduced by Microsoft Research (2024), it targets questions that require *connecting* facts across many documents — multi-hop and whole-corpus queries — that similarity-only chunk retrieval cannot answer. ## Why chunk retrieval falls short - Standard RAG retrieves independent [[RAG Chunking Strategy|chunks]] by [[Embedding-Based Retrieval|vector similarity]] — excellent for "find the passage," poor for "how are these related?" or "summarize the themes across everything." - Disconnected chunks carry no explicit relationships between the facts they contain. ## How it works 1. **Extraction** — an LLM reads the chunks and extracts entities and relationships, assembling a graph. 2. **Community detection** — cluster the graph into communities and summarize each. 3. **Query** — *local search* (an entity's neighbourhood) for specific questions; *global search* (community summaries) for corpus-wide questions. ## Trade-off More powerful for connected and global questions, but the graph is **expensive to build and maintain** (LLM extraction over the whole corpus) compared with plain [[Vector Database|vector]] RAG. Reach for it when the relationships *between* facts are the point. ## Related - [[Retrieval-Augmented Generation]] - [[RAG Chunking Strategy]] - [[Hybrid Search]] - [[Embedding-Based Retrieval]] - [[Graph Neural Network]] ## Sources - Edge, D., et al. (2024). *From Local to Global: A Graph RAG Approach to Query-Focused Summarization.* Microsoft Research. arXiv:2404.16130.