Migrate a vector database to a new embedding model by fitting a transform from ~2K calibration texts. 86–92% retrieval retention on BEIR.
Incompatible spaces make cross-model search fail. Re-embedding is the expensive fix — and deprecation keeps forcing the cycle.
Different models produce vectors in different geometric spaces. Raw cross-space nDCG@10 = 0.0.
1024d from bge ≠ 1024d from e5
Switching models means re-embedding every document. A million docs means hundreds of dollars — and downtime while you rebuild.
Cost scales linearly with corpus size
OpenAI deprecated ada-002. Newer models are better, but migrating still means rebuilding from scratch — unless you learn the map.
ada-002 → text-embedding-3: same story
Fits a ridge regression from ~2K paired calibration texts. The learned mapping transforms stored vectors, or maps new-model queries into the legacy space.