NDCG

Logs behind a fluent RAG demo showed the correct chunk arriving 30% of the time

A generator writes with equal confidence whether its context is right or wrong, so the only readout that separates the layers is a labelled eval set you own, scored on the retrieval side before an answer exists.