USA
Canada
UK
Europe
Australia
Singapore
Japan
IndiaQ1. Walk through the components of a RAG pipeline.
Short answer: Ingest → chunk → embed → store → retrieve → (re-rank) → generate with grounding.
Strong answer: A RAG pipeline ingests documents, chunks them thoughtfully, embeds chunks, stores them in a vector DB with metadata, retrieves top-k relevant chunks for a query (often hybrid keyword+vector), optionally re-ranks, then prompts the LLM with that context and citations. Evaluation and guardrails wrap it. Retrieval quality drives most of the outcome.
Likely follow-up: Which step most often causes poor RAG answers?
What the interviewer is checking: End-to-end RAG understanding.
Common mistake: Focusing on the LLM while ignoring retrieval quality.
Q2. How do you choose a chunking strategy?
Short answer: Balance chunk size and overlap to preserve meaning without diluting retrieval.
Strong answer: Chunk by semantic boundaries (paragraphs/sections) with a size that fits context and keeps ideas intact, plus some overlap so context isn’t cut mid-thought. Too large dilutes relevance and wastes tokens; too small loses context. Structure-aware chunking (headings, tables) and metadata improve retrieval.
Likely follow-up: What problem does chunk overlap solve?
What the interviewer is checking: Retrieval quality knowledge.
Common mistake: Fixed tiny or huge chunks with no regard for structure.
Q3. What are embeddings and how are they used in RAG?
Short answer: Vector representations of meaning used to find semantically similar chunks.
Strong answer: Embeddings map text to vectors where similar meanings are close. In RAG you embed chunks and the query, then find nearest neighbors in a vector DB as candidate context. Choice of embedding model and dimensionality affects retrieval quality and cost; domain-specific embeddings can help.
Likely follow-up: Why might two embedding models give different retrieval quality?
What the interviewer is checking: Core RAG plumbing.
Common mistake: Confusing embeddings with the generative model.
Q4. What is hybrid search and why use it?
Short answer: Combining keyword (lexical) and vector (semantic) retrieval for better recall.
Strong answer: Hybrid search blends lexical (BM25/keyword) and semantic (vector) retrieval, then fuses the results. It catches exact terms/codes that embeddings miss and semantic matches that keywords miss — improving recall and robustness, especially for names, IDs and rare terms. Re-ranking on top further boosts precision.
Likely follow-up: When does pure vector search fail?
What the interviewer is checking: Retrieval design depth.
Common mistake: Assuming vector search alone is always best.
Q5. How do you evaluate RAG quality?
Short answer: Measure retrieval relevance and answer groundedness/faithfulness against references.
Strong answer: Evaluate retrieval (are the right chunks returned? recall/precision@k) and generation (is the answer grounded in retrieved context and correct? faithfulness, answer relevance). Use labeled eval sets, LLM-as-judge with care, and human review. Separating retrieval vs generation failures tells you what to fix.
Likely follow-up: How do you tell a retrieval failure from a generation failure?
What the interviewer is checking: RAG evaluation rigor.
Common mistake: Only eyeballing final answers without measuring retrieval.
Q6. How do you keep RAG answers grounded and reduce hallucination?
Short answer: Constrain to retrieved context, cite sources, and instruct abstention.
Strong answer: Prompt the model to answer only from retrieved context, cite the source chunks, and say it cannot find the answer when context is missing rather than inventing one. Add relevance thresholds so weak retrievals don’t force a guess, and verify high-stakes answers. Grounding + citations + abstention is the core defense.
Likely follow-up: What do you do when retrieval returns irrelevant chunks?
What the interviewer is checking: Grounding strategy.
Common mistake: Letting the model answer freely regardless of retrieval quality.
Q7. How do you keep a RAG knowledge base fresh?
Short answer: Incremental ingestion, re-embedding on change, and metadata for recency.
Strong answer: Set up incremental pipelines that ingest new/changed documents, re-embed and upsert them, and remove stale content, with metadata (timestamps, source) to filter or prefer recent data. Schedule or event-trigger updates so the index reflects current reality. Stale indexes are a common cause of wrong answers.
Likely follow-up: How would you handle a document that was deleted at the source?
What the interviewer is checking: Operational RAG.
Common mistake: Building the index once and never refreshing it.