AI Vector Database Engineer Interview Questions

A great answer explains dense numerical representations of data, semantic similarity via distance metrics, and how embeddings capture meaning beyond keyword matching.

Cover when each metric is appropriate, how cosine is magnitude-invariant, and how dot product correlates with cosine for normalized vectors.

Discuss token limits, retrieval granularity, and strategies like fixed-size, recursive character splitting, and semantic chunking.

Cover specialized index structures (HNSW, IVF), similarity search vs. exact match, approximate nearest neighbor tradeoffs, and metadata filtering capabilities.

Explain it as a PostgreSQL extension for vector operations, its advantages for teams already on Postgres, and its limitations at very large scale.

Cover the multi-layer graph structure, how M controls connectivity, ef_construction affects build quality, ef_search affects query accuracy-latency tradeoff.

Discuss bucket-based partitioning, nprobe parameter for search breadth, PQ for memory compression, and suitability for very large datasets with lower memory budgets.

Cover reciprocal rank fusion (RRF), weighted score normalization, two-stage retrieval with re-ranking, and tools like Weaviate's hybrid search or Elasticsearch kNN with text query.

Discuss recall@k, precision@k, MRR, NDCG, latency percentiles (p50, p95, p99), and the importance of a golden evaluation dataset with ground-truth relevance labels.

Cover pre-filtering vs. post-filtering tradeoffs, index integration with metadata, and how different databases (Pinecone, Qdrant, Weaviate) handle this differently.

Discuss collection aliasing, blue-green index deployments, embedding version metadata, and backward-compatible migration strategies.

Cover subspace decomposition, codebook training, lossy compression, and the recall degradation vs. memory savings curve.

Discuss tenant-based partitioning, metadata-based filtering, namespace/collection strategies, and row-level security approaches.

Explain two-stage retrieval, cross-encoder re-rankers (e.g., Cohere Rerank, bge-reranker), and the latency-accuracy tradeoff of adding a re-ranking step.

Cover TCO analysis, operational overhead, scalability requirements, data residency constraints, and feature maturity comparisons.

Cover sharding strategy, HNSW vs. IVF-PQ choice, memory vs. disk-based tiers, replica configuration, and DNS-level load balancing.

Discuss micro-batch upserts, idempotency keys, schema evolution handling, eventual consistency guarantees, and backpressure mechanisms.

Cover the curse of dimensionality, distance concentration, the 'hubness' problem, and mitigation strategies like dimensionality reduction or graph-based indices.

Discuss monitoring embedding norms, pairwise distance distributions, retrieval recall drift, automated alerting, and re-embedding pipelines.

Cover shadow indexing, dual-collection serving, traffic splitting at the query layer, and statistical significance evaluation of retrieval quality metrics.

Discuss WAL-based backups, snapshot strategies, cross-region replication lag, RTO/RPO targets, and automated failover with health checks.

Discuss shared embedding space, separate collections vs. unified index, query routing, and cross-modal re-ranking strategies.

Cover memory savings, recall degradation curves, Matryoshka representations, and when quantization is acceptable vs. when precision is critical.

Discuss pre-filtering with ACL metadata, post-retrieval filtering, separate namespaces, and the security implications of embedding information leakage.

Cover vector size × count memory estimation, index build time projections, shard rebalancing strategies, and cost modeling across managed vs. self-hosted options.

Cover embedding dimension mismatch, re-indexing requirements, evaluation against golden dataset, and rollback procedures.

Cover query profiling, index memory pressure, HNSW ef_search tuning, shard hotspots, and horizontal scaling triggers.

Discuss HNSW non-determinism, index rebuild differences, floating-point precision across hardware, and ensuring identical index build parameters.

Cover dual-write pattern, incremental sync, cutover validation with shadow traffic, rollback plan, and schema mapping between platforms.

Discuss tombstone deletion, collection-level vs. vector-level deletion, index compaction, and audit logging for deletion verification.

Cover retrieval evaluation (recall@k, MRR on ground truth), context window analysis, prompt engineering for grounded generation, and attribution tracing.

Discuss multi-modal embeddings, weighted score fusion, personalized re-ranking, and the tradeoffs of unified vs. separate indices per modality.

Cover benchmark dataset preparation, latency/recall testing under load, operational features (backup, replication, auth), community/support maturity, and TCO.

Discuss encryption at rest and in transit, de-identified embeddings, VPC isolation, audit logging, access control, and BAA requirements with managed service providers.

Cover semantic chunking, minimum chunk length enforcement, overlap optimization, quality scoring heuristics, and re-indexing strategy.

Cover Qdrant client configuration, gRPC vs. REST API selection, retry logic, connection pooling in async contexts, and LangChain's VectorStore interface.

Cover document loaders, node parsers, embedding model configuration, Weaviate vector store integration, query engine setup, and response synthesis.

Discuss infrastructure-as-code for vector DB provisioning, migration scripts for collection schema changes, embedding pipeline versioning, and automated benchmark gates.

Cover model loading, batch encoding with device management, Pinecone batch upsert limits, metadata attachment, and progress tracking for long-running jobs.

Discuss query latency histograms, index memory usage, collection size growth, error rates, and alert thresholds for latency degradation and capacity planning.

Cover OpenSearch kNN field configuration, hybrid query DSL with bool and knn clauses, score combination strategies, and relevance tuning.

Discuss Cohere Embed for ingestion, Qdrant for storage and initial retrieval, Cohere Rerank as a second stage, and latency optimization.

Cover namespace-level isolation guarantees, metadata filtering within namespaces, index-level isolation for compliance, and cost implications of each approach.

Discuss context injection into system/user prompts, source document metadata for citations, token budget management, and response parsing for attribution.

Cover Kafka topic design, consumer group configuration, embedding computation in the stream, Milvus batch upsert API, dead-letter queues, and exactly-once semantics.

A great answer shows structured decision-making, stakeholder communication, data-driven benchmarking, and the outcome of the chosen tradeoff.

Look for analogies (library catalog, GPS coordinates), awareness of audience, and the ability to connect technical concepts to business outcomes.

A strong answer covers incident response process, root cause analysis, communication with stakeholders, and concrete post-mortem improvements.

Look for specific sources (research papers, community forums, benchmarks), proactive experimentation, and how learning translated to tangible improvements.

A great answer demonstrates principled technical leadership, data-driven persuasion, and balancing business pressure with engineering quality.