Vector Database Interview Q&A
Core Questions
Q: What is a vector database?
A: A database optimized to store embeddings and answer nearest-neighbour queries fast, using ANN indexes and metadata filtering.
Q: Exact vs ANN search?
A: Exact search compares the query against every vector and returns perfect results. ANN trades slight recall loss for orders-of-magnitude speed using structures like HNSW, IVF or quantization.
Q: Cosine, dot product or L2?
A: Cosine for meaning/text, dot product for normalized vectors or magnitude-aware ranking, L2 for geometric features. Match metric between embedding, index and query.
Indexing Questions
Q: How does HNSW work?
A: It builds a multi-layer graph. Higher layers link long-range neighbours for fast navigation; lower layers refine the result. M and efSearch control build quality and query recall.
Q: What is product quantization?
A: Each vector is split into sub-vectors that map to short codes, shrinking memory dramatically in exchange for approximating distances.
Practical Questions
Q: How do you update stale vectors?
A: Content-hash the ids, upsert on change, delete by filter when a source is removed, and re-run the eval set to catch drift.
Q: How do you combine vectors with filters?
A: Pre-filter when the filter is selective; post-filter when it is broad. Store payload metadata on every point.
Q: How do you benchmark?
A: Fixed question-to-answer set, recall@k and precision@k, latency percentiles, and memory usage at real data scale.
Key Points
- Explain ANN tradeoffs with numbers, not slogans.
- Keep the metric, index and query in agreement.
- Cite recall, latency and memory together.
- Discuss updates and monitoring, not just search.