RAG's top-k retrieval has a black box problem, and it isn't fixed by a bigger model

A vector search returns the nearest chunks by embedding distance, hands them to the model, and stops. Nobody asks why those chunks and not the ones one rank lower. When retrieval brings back the wrong