Course Content
Scenario-Based AI Engineering Questions
26 sections · 146 lessons
How do you handle the probing follow-up questions an interviewer asks after your first answer?
What you need to know
What probes are for
Probing questions are the follow-ups after your first answer: "Why not just increase top-K?", "What does that cost?", "What happens when it fails?" Your first answer shows you have read about the topic. The probes show whether you have built it. They test two things: how deep your understanding goes, and whether you know where it stops. Confidently inventing a number is worse than saying you would measure it.
The depth ladder
Each probe usually asks you to go one level down. Know what the next level is before you are asked.
| Level | The question behind it | Example for "use a cross-encoder reranker" |
|---|---|---|
| Design | What would you build? | "Retrieve 50 with hybrid search, rerank to 5." |
| Mechanism | How does it work, and why does it help? | "It reads query and passage together, so it catches meaning a bi-encoder misses." |
| Numbers | What does it cost, and how would you know it worked? | "It adds latency for every candidate, so I'd rerank 50, not 500, and track recall@5 and p95." |
| Failure modes | When does it break, and what then? | "Long passages get truncated; I'd rerank chunks, not whole documents." |
How to handle a probe
- Answer the question asked — if they ask about top-K, talk about top-K. Rephrasing your first answer is the most common failure.
- Go one level deeper — design, then mechanism, then numbers, then failure modes.
- Name the trade-off — "This adds a second model to run. I'd accept that for legal search, not for autocomplete."
- Say what you'd measure — "I don't know the exact figure; I'd A/B it and watch recall@10 and p95 latency."
- Concede fast when they're right — "Good point, that breaks for multi-tenant data. I'd move the filter into the index." Then continue with the better design.
Phrases that work
- "The trade-off is ___; I'd choose it because ___, unless ___."
- "I haven't run that at this scale. My estimate is ___, and I'd confirm it by ___."
- "You're right, that fails when ___. I'd change it to ___."
A real-life example
A mock interview, made up. The question is a RAG system that returns wrong answers. The candidate answers well: add hybrid search and a reranker. The interviewer probes: "Why not just increase top-K from 5 to 20?"
A weak reply repeats the first answer: "Because a reranker improves relevance." A strong reply answers the probe: "More chunks raise recall but put more noise into the prompt, and cost grows with every token. In our last project, going from 5 to 20 chunks raised cost per query about 3 times and did not improve answer accuracy, because the right chunk was often at rank 30. The fix was better ranking, not more slots." The interviewer then asks for the reranker's latency. The candidate says: "I don't remember the exact figure for that model. I'd benchmark it on our p95 budget, and if it's too slow I'd rerank fewer candidates or use a smaller reranker." That honest answer, with a clear way to find out, scores better than a guessed number.
Follow-up questions to expect
- "What if I genuinely don't know the answer?" — Say so briefly, then reason out loud from what you do know and say how you'd find out. Interviewers value the reasoning more than the fact.
- "What if the interviewer disagrees with a correct answer?" — Ask what case they have in mind. They may be testing whether you hold a correct position; explain your reasoning calmly, and change only if they show a real gap.
- "How long should each answer be?" — Short: a direct answer in two or three sentences, then stop and let them choose where to probe next.