Embedding drift
The problem that embeddings from different models (or model versions) aren't comparable — so when you switch or update the embedding model, the entire vector database must be recomputed. A notorious maintenance problem in production RAG.
Mechanic: a vector search only works if the query and documents lie in the same embedding space. Different models (or a new version of the same) place text at entirely different coordinates → a query embedded with model B can't be meaningfully compared with documents embedded with model A. Consequence: if you want to upgrade to a better embedding model you must re-embed and re-index EVERYTHING, which for large corpora is expensive and time-consuming. The term also covers the data's content/distribution sliding over time so old embeddings become less representative. Mitigation: matryoshka embeddings and version management, plus planning for re-indexing. A practical reality every RAG system in operation faces. Related to embedding model and matryoshka embedding.