Drift – an embedding-model upgrade should be a rotation, not a reindex

Drift proposes that upgrading an AI embedding model should be treated as a learned rotation of the existing vector space rather than a full reindexing of all data. This technique allows developers and ML engineers to seamlessly adopt new, more powerful models without the massive computational cost and downtime of regenerating embeddings for entire databases. It is interesting because it solves a critical scalability bottleneck in retrieval-augmented generation (RAG) and vector search, making model iteration practical for production systems.