Retrieval-Augmented Generation (RAG) Interview Questions Interview Guide

Sample Answer:

Built RAG system for internal documentation. Tracked: retrieval precision (did top-K contain answer?), answer accuracy (human evaluation), latency (<500ms target), and cost per query. Achieved 88% accuracy with continuous improvements.

Sample Answer:

Initial setup: vector search with text-embedding-ada-002. Precision was 70%. Iterated: tried larger embedding model (78%), added re-ranking with LLM (85%), implemented hybrid search (89%). Final system: hybrid + re-ranking.

Sample Answer:

System started hallucinating answers not in documents. Root cause: prompt didn&apos;t instruct model to use only retrieved context. Added explicit instruction "Answer only from retrieved documents." and confidence filtering. Hallucinations dropped 95%.

Sample Answer:

Test chunk sizes: 256, 512, 1024 tokens. Measure retrieval precision and latency. Default 512 with 128 overlap balances context and relevance. Use semantic chunking for complex documents.

Sample Answer:

Naive: retrieve top-K, pass to LLM. Advanced: add re-ranking, query expansion, multi-step retrieval, parent-child retrieval, hypothetical questions. Implement each, measure improvement, stack highest-impact techniques.

Sample Answer:

Chunk into semantic sections, embed with context windows, implement hierarchical retrieval (find section first, then passage), use summarization for background context. Pre-process documents into structured format.

Sample Answer:

Use RAGAS metrics: context precision, context recall, answer relevance. Run human evaluation on sample. Track drift weekly. Implement A/B testing for improvements. Monitor confusion cases.