Interview Q&A · 22 questions · 4 answered publicly
Vector Databases Interview Questions
Ordered by how often they're asked. Sign in free to see every answer, rate your confidence and run timed mock interviews.
-
Most asked Vector DB leaf: What is an embedding?
Summary
A dense vector representing semantic content so similar items are close under a chosen metric (cosine/dot/L2).
Longer Answer
A dense vector representing semantic content so similar items are close under a chosen metric (cosine/dot/L2). Foundation.
Key Points
- A dense vector representing semantic content so similar items are close under a chosen metric (cosine/dot/L2)
- Foundation
- Relevant areas: Ai, Vector Db, Embeddings
- Use the answer to connect the concept back to a realistic production or interview scenario for 'Vector DB leaf: What is an embedding?'
Comparison / Tradeoffs
Aspect Notes Core idea A dense vector representing semantic content so similar items are close under a chosen metric (cosine/dot/L2). Practical tradeoff Foundation. -
Most asked Vector DB leaf: HNSW vs IVF indexes?
Summary
HNSW: graph ANN, strong recall/latency, higher memory.
Longer Answer
HNSW: graph ANN, strong recall/latency, higher memory. IVF: cluster-based, tunable nprobe, often cheaper at huge scale.
Key Points
- HNSW: graph ANN, strong recall/latency, higher memory
- IVF: cluster-based, tunable nprobe, often cheaper at huge scale
- Index tradeoffs
- Use the answer to connect the concept back to a realistic production or interview scenario for 'Vector DB leaf: HNSW vs IVF indexes?'
Comparison / Tradeoffs
Aspect Notes Core idea HNSW: graph ANN, strong recall/latency, higher memory. Practical tradeoff Index tradeoffs. -
Most asked Vector DB leaf: When is hybrid search necessary?
Summary
Exact IDs, rare tokens, and code-combine BM25 with vectors and fuse scores (e.g., RRF).
Longer Answer
Exact IDs, rare tokens, and code-combine BM25 with vectors and fuse scores (e.g., RRF). Retrieval quality.
Key Points
- Exact IDs, rare tokens, and code-combine BM25 with vectors and fuse scores (e.g., RRF)
- Retrieval quality
- Relevant areas: Ai, Vector Db, Hybrid Search
- Use the answer to connect the concept back to a realistic production or interview scenario for 'Vector DB leaf: When is hybrid search necessary?'
Comparison / Tradeoffs
Aspect Notes Core idea Exact IDs, rare tokens, and code-combine BM25 with vectors and fuse scores (e.g., RRF). Practical tradeoff Retrieval quality. -
Most asked Vector DB leaf: Cosine vs dot product-when does it matter?
Summary
If embeddings are L2-normalized, cosine and dot align.
Longer Answer
If embeddings are L2-normalized, cosine and dot align. Otherwise choose metric matching training assumptions.
Key Points
- If embeddings are L2-normalized, cosine and dot align
- Otherwise choose metric matching training assumptions
- Similarity metrics
- Use the answer to connect the concept back to a realistic production or interview scenario for 'Vector DB leaf: Cosine vs dot product-when does it matter?'
Comparison / Tradeoffs
Aspect Notes Core idea If embeddings are L2-normalized, cosine and dot align. Practical tradeoff Similarity metrics. - Vector DB leaf: How do you filter by metadata efficiently?
- Vector DB leaf: Re-ranking after ANN retrieval?
- Vector DB leaf: What causes embedding drift?
- Vector DB leaf: Dimensionality vs quality/cost?
- Vector DB leaf: How do you delete/update vectors safely?
- Vector DB leaf: Multi-tenancy strategies?
- Vector DB leaf: How do you benchmark a vector DB?
- Vector DB leaf: Sparse + dense vector fusion?
- Vector DB leaf: What SLIs matter for vector search?
- Vector DB leaf: Explain product quantization briefly.
- Vector DB leaf: What is a collection/namespace?
- Vector DB leaf: How do you store vectors for multimodal data?
- Vector DB leaf: Cold start for a new corpus?
- Vector DB leaf: Why normalize embeddings?
- Vector DB leaf: Handling near-duplicate documents?
- Vector DB leaf: When to choose managed vs self-hosted vector DB?
- What are embeddings, and how does vector similarity search find relevant documents?
- When would you use hybrid search (BM25 + vectors) instead of pure vector search?