RAG Interview Questions and Answers
Core Questions
Q: What is RAG?
A: Retrieval-Augmented Generation fetches relevant documents first, then asks the model to answer using them, grounding answers in real evidence.
Q: How does RAG reduce hallucinations?
A: The prompt restricts the model to retrieved notes, and the model is allowed to say it doesn't know. Answers trace back to citable chunks.
Q: RAG vs fine-tuning?
A: RAG supplies fresh facts without training; fine-tuning changes behaviour and style. Most systems use both, for different jobs.
Design Questions
Q: How do you choose a chunk size?
A>... A small-to-medium chunk with overlap and structure-aware boundaries balances precision and context, then validate against your eval set.
Q: What is hybrid search?
A: Running dense vector retrieval plus keyword retrieval (BM25) and fusing the ranked lists, usually with reciprocal rank fusion.
Q: How do you evaluate a RAG system?
A: Context relevance, faithfulness and answer relevance, scored against a fixed question set, plus latency and cache metrics online.
Scenario Questions
Q: Your knowledge base changes daily.
A: Content-hash chunk ids, re-index changed sources, and re-run the eval set each release.
Q: Answers mix facts from several files.
A: Retrieve wider, rerank, and prompt for multi-source synthesis with citations per source.
Key Points
- Answer with the pipeline and its tradeoffs.
- Quote concrete metrics and eval methods.
- Show you handle updates, cost and drift.
- Mention grounding prompts with "I don't know".