When you change embedding models, the stored vectors should be treated as belonging to a different embedding space unless you intentionally keep the entire corpus consistent. Microsoft’s vector guidance notes that when most or all embeddings are replaced with fresh embeddings from a new model, the recommended practice is to reload the new embeddings and, for large-scale replacement scenarios, consider dropping and recreating the vector index afterward so search quality remains predictable.

This question also says applications must continue to use VECTOR_SEARCH without runtime errors . VECTOR_SEARCH requires compatible vector dimensions, and the vector column already exists. Azure OpenAI documentation shows that text-embedding-ada-002 is fixed at 1536 dimensions and text-embedding-3-small supports up to 1536 dimensions . That means the migration can remain compatible with a VECTOR(1536) column, but the right implementation step is still to re-embed the existing rows so the table does not contain a mixed corpus produced by different models.

The other options are not appropriate:

B normalization does not solve a model migration problem.

C converting the vector column to nvarchar(max) would break vector-native search design.

D a vector index improves performance, but it does not migrate old embeddings to the new model.