Your Embedding Model Will Deprecate. Here's What to Do.

- Embedding model providers (OpenAI, Cohere, Google, AWS) deprecate older models on a regular cadence. When it happens, every vector in your index needs to be regenerated.- Embeddings from different models are geometrically incompatible, even when dimensions match. There is no shortcut: you have to re-embed.- Three production strategies: blue-green index deployment (build a parallel index and cut over), mixed-model indexes with RRF fusion (migrate gradually while keeping both queryable), and embedding space alignment (promising research, but no confirmed production deployments yet).- Standard A/B testing is misleading for embedding swaps because the retrieval step itself changes. Use LLM-as-judge for offline validation and canary rollouts with automated rollback.- Build for migration from day one: version your embeddings, store the original text alongside the vectors, and keep a retrieval evaluation harness ready. Teams that treat the embedding model as a permanent decision scramble when the deprecation notice arrives.