My lab found a way to migrate between embedding models with zero downtime. [R]

A lab has discovered a method for migrating between embedding models with zero downtime. This innovation addresses the significant challenge of upgrading embedding models, which can take considerable time, especially with large datasets. For instance, re-embedding a billion documents on an H100 GPU could take 108 days. The new method aims to allow users to upgrade models without service interruption, even for datasets as small as 50 million vectors.