RAG Architecture Evolution

GraphRAG: when flat-chunk retrieval hits 25.7% and what replacing it costs

Flat-chunk RAG answers temporal multi-hop questions at 25.7% accuracy — a floor the architecture cannot escape. This issue dissects what GraphRAG adds to fix that, what it costs in indexing ($50–$200 vs. under $5 for a 500-page corpus), where it still loses to plain RAG, and the exact conditions under which neither approach makes sense.