Embedding Drift

Embedding drift is what happens when the meaning or distribution of your vector embeddings shifts over time, degrading a retrieval or semantic-search system. It comes in two flavors. Model drift: you upgrade or swap the embedding model, and the new vectors live in a different space — incompatible with everything already indexed, so you must re-embed your whole corpus. Data drift: the content or the queries your users send gradually change (new products, new jargon, new topics), so an index built last year retrieves poorly today. For SaaS builders running RAG or semantic search, drift is a silent quality killer — relevance slowly rots while nothing throws an error. The defenses: never mix embeddings from different models in one index, version your embedding model alongside your data, re-embed when you upgrade, and monitor retrieval quality with a fixed eval set so you notice degradation early. Budget for periodic re-indexing; embeddings are an ongoing cost, not a one-time build.