RAG Interview Questions, issue 4, Jul 8, 2026
The Parametric Memory Trap
Staff ML Engineer interview at Google, and the interviewer asks:
“We want to swap our BM25 + dense pipeline for generative retrieval. Our corpus gains 50,000 documents every Monday. Why might that be a dealbreaker?”
Don’t say: “You just re-index the new docs like any other system.”
The hidden architectural flaw where swapping BM25 for generative retrieval stops saving time and starts destroying your old data just to learn new document IDs.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.