Interview prep
RAG interview questions with answers: 15 you should be ready for
Fifteen retrieval-augmented generation interview questions with short answers: chunking, embeddings, reranking, hybrid search, evaluation and failure modes.
Short answer: interviewers test whether you understand each stage of a RAG system (chunk, embed, retrieve, rerank, generate) and how you measure it. Learn the stages first, then the failure modes. Background: RAG explained.
Fundamentals
- What is RAG and why use it? It retrieves relevant text and gives it to the model with the question, so answers use your data and can cite sources, without retraining.
- RAG or fine-tuning? RAG for facts that change or must be cited; fine-tuning for style or a fixed behaviour. See RAG vs fine-tuning.
- What is an embedding? A vector that places text so that similar meaning is close. See embeddings explained.
- What is cosine similarity? The angle-based closeness of two vectors; the usual way to rank retrieved chunks.
Building the pipeline
- How do you choose a chunk size? Test it. Too small loses context, too large adds noise. Overlap helps at the edges. See context fragmentation.
- What is hybrid search? Combining keyword and vector search, then merging the results. See hybrid search.
- What does a reranker do? It re-scores the top retrieved chunks with a stronger model so the best ones go first.
- How do you handle tables, PDFs and code? Parse them so structure survives; do not flatten everything to plain text.
- How do you keep the index fresh? Re-index changed documents on a schedule or on events, and store a version or date.
Quality and safety
- How do you evaluate a RAG system? Measure retrieval (did the right chunk come back?) and the answer (is it correct and supported by the chunks?) on a labelled set. See LLM evaluation pipeline.
- What is hallucination in RAG and how do you reduce it? The model states something the context does not support. Require citations, say "I do not know" when context is empty, and check answers against sources.
- What is prompt injection? Text in a document that tries to give the model new instructions. Treat retrieved text as data and add guardrails.
- How do you control cost and latency? Cache, retrieve fewer chunks, use smaller models where they are enough.
Design questions
- Design a question-answering system over company documents. Cover ingestion, permissions (who may see what), retrieval, generation, evaluation, monitoring and feedback.
- Your answers are wrong. What do you check first? Look at the retrieved chunks. If they are wrong, fix retrieval; if they are right but the answer is wrong, fix the prompt or the model.
Practise
Build one RAG project and evaluate it (RAG portfolio projects). The Synaptova program teaches this in its retrieval module: see the program, or take the free AI Readiness Test.