RAG Tutor: Changing the Embedding Model on a Live System

This retrieval-augmented tutor finds source passages before it answers. Each answer cites its source passage.

An embedding model turns text into a vector: a list of numbers. Each number is one dimension. Our vector index stores vectors for search, up to a dimension limit. The larger model we wanted went over that limit. We used it at reduced dimensions instead of a smaller model.

The live migration used a shadow column for the new vectors and a second search function. We wrote the rollback script before the migration ran.