1Cademy - GraphRAG Framework (Edge et al., 2024)

Learn Before

Research

Concept

GraphRAG Framework (Edge et al., 2024)

GraphRAG is a graph-based retrieval-augmented generation framework introduced by Edge et al. (Microsoft Research, 2024) for query-focused summarization over an entire corpus. Indexing has two stages: (i) an LLM extracts an entity-and-relationship knowledge graph from the source documents, with element-level descriptions; and (ii) hierarchical community detection (Leiden algorithm) partitions the entity graph into nested clusters, and an LLM pre-generates a community summary for every community at every level. At query time, GraphRAG operates in a map-reduce fashion over a selected community level: each relevant community summary independently produces a partial answer (map), and the partial answers are aggregated into a final response (reduce). The framework is explicitly designed to handle global sensemaking questions over a full corpus that flat dense RAG cannot answer well, and is the canonical reference for any 'GraphRAG-style' graph retriever baseline in subsequent RAG evaluations.