These are RAG questions of the kind Google actually asks — the patterns reported from Google's AI-engineering rounds, where design trade-offs, scale and failure modes matter as much as definitions. Treat this page as a mock interview: say every answer out loud before revealing it. If one surprises you, the lesson behind it is linked at the bottom.
Google RAG concept questions
With million-token context windows, make the strongest engineering case that retrieval is still necessary — and name the case where it isn't.
Asked in
You use an LLM to judge whether RAG answers are grounded. What can go wrong with the judge itself, and how do you control for it?
Asked in
Explain HyDE and why it can improve retrieval even when the hypothetical answer it generates is factually wrong.
Asked in
Google RAG applied & hands-on questions
Three RAG systems report these numbers. For each, say what is wrong and what you'd fix first. A: recall@5 = 45%, faithfulness = 0.94. B: recall@5 = 96%, faithfulness = 0.55. C: recall@5 = 93%, faithfulness = 0.91, refusal correctness = 0.30.
Asked in
How to use this page: Google rarely asks something you've never seen — they ask a standard RAG concept and then push one level deeper ("why?", "what would you do if..."). Master the concept in the RAG course lessons, and the follow-up stops being scary.
Keep practising: What Is RAG?, Chunking, Retrieval Techniques and Evaluating RAG cover what most Google RAG rounds test.