Vector Databases & Semantic Search Interview Questions Interview Guide

Sample Answer:

Chose Weaviate over Pinecone for our RAG system. Evaluated: cost (self-hosted Weaviate saved 70%), query performance (sub-100ms required), filtering capabilities (complex metadata queries), and operational overhead. Weaviate won on cost and features.

Sample Answer:

Started with single node, hit QPS limits at 500 requests/sec. Implemented sharding, added replicas, optimized index parameters (HNSW searcher depth), and achieved 10K QPS with <50ms latency.

Sample Answer:

Changed embedding model from text-embedding-ada-002 to BGE-large for better domain fit. Reindexed 2M vectors in background, ran parallel searches during migration, cutover was seamless.

Sample Answer:

Embedding dimension affects storage and speed (768 vs 1536). Model quality affects search relevance. Choose embedding based on domain (general vs specialized), dimension trade-off (size vs accuracy), update frequency.

Sample Answer:

HNSW builds hierarchical graphs, offers fast search and better precision, uses more memory. IVF partitions space into clusters, faster indexing, lower memory but lower precision. Choose HNSW for <1M vectors, IVF for 1B+ vectors.

Sample Answer:

Execute vector search and BM25 keyword search in parallel, weight results (e.g., 0.6 vector + 0.4 keyword), re-rank combined results. Use reciprocal rank fusion for fair weighting.

Sample Answer:

Implement versioning for embeddings, batch update stale vectors nightly, monitor query drift with human feedback, retrain embedding model quarterly. Keep audit trail of vector changes.