150 interview questions, 103 traced to candidate and vendor reports, with senior answers, for 2027 loops
RAG is now the most common GenAI design question in big tech interviews. Most prep material stops at "put documents in a vector database and call a model." Senior interviewers expect much more. They ask how you keep permissions correct, how you know a change helped, and what breaks when real traffic arrives.
We wrote this guide for that interview.
What is inside
A question bank built from real reports. 150 RAG interview questions, organized by company: Google, Meta, Amazon, Microsoft, Apple, OpenAI, Anthropic, NVIDIA, Databricks, Glean, Perplexity, Cohere, Harvey and more. 103 questions trace to candidate reports on Glassdoor, 1point3acres and Exponent, or to prep sites that tag them to a company. Every source is listed, with a date, so you can judge how current it is.
Senior answers, not keyword lists. Each answer gives the opening you say in the first two minutes. Then it gives the design, the trade-offs and the capacity math. Each card also lists the follow-up probes interviewers ask next, and two self-check lines: what a strong answer must include, and the most common miss.
Ten teaching chapters. Part one teaches the material the bank tests:
- RAG foundations and system design
- Document ingestion, parsing and chunking
- Embeddings and vector databases
- A complete basic RAG system with citations
- Hybrid retrieval and reranking
- Evaluation and observability
- Knowledge graph fundamentals
- GraphRAG and hybrid graph retrieval
- Agentic GraphRAG and multi-hop reasoning
- Production design: latency, cost, security and scale