Embeddings in Production: Versioning, Drift, and Retrieval Quality

Embedding interviews focus on the retrieval contract, not vector dimension trivia. Interviewers ask how you choose embedding models for legal versus conversational text, whether query and document embeddings should be asymmetric, and how normalization affects cosine similarity rankings in production indexes.

Embedding drift is a leading cause of silent RAG regressions. Model upgrades, tokenizer changes, and domain shift all require re-embedding strategies. Strong candidates describe dual-index migrations, shadow embedding comparisons, and rollback plans when recall drops on critical slices after a model swap.

Dimensionality and latency trade-offs matter at scale. Higher-dimensional embeddings can improve recall but increase index size and search cost. Discuss when matryoshka embeddings help, when PCA or quantization is acceptable, and how to validate that compression does not break tail-query performance.

Evaluation for embeddings should use task-specific benchmarks, not generic STS scores. Interviewers expect discussion of hard negatives, in-domain query-document pairs, and slice analysis for multilingual or acronym-heavy corpora. Offline recall@k should correlate with downstream answer faithfulness.

Operational concerns include batch versus online embedding, deduplication of near-duplicate chunks, and caching embeddings for stable documents. Articulate how you would detect when two chunks embed too similarly and collapse retrieval diversity—a common failure in dense FAQ corpora.