Notes on RAG evaluation

24 Oct 2026 · RAG evaluation

I said GraphRAG beats vector RAG. Then I measured it.

A small pre-registered experiment: when a graph retriever and a tuned hybrid retriever get the same context budget, how much of the graph advantage is left? +19.0 F1 points on 2WikiMultiHopQA, +3.5 on MuSiQue, with the limits stated.