Why basic RAG fails at multi-hop reasoning (and how GraphRAG fixes it)

The New Stack · 1 month ago · 21

The article explains that standard chunk-based RAG often fails on multi-hop and global summary questions because semantically similar chunks don’t reliably contain the needed linked entities. It points to a concrete example where splitting contracts into 1,000-token chunks causes a query about who leads an acquired company to miss the CEO details. It proposes GraphRAG, which ingests documents into a knowledge graph and then uses graph traversal (via a retriever) so the system can retrieve connected entities needed for multi-hop reasoning.