What is embedding alignment evaluation across model versions?

Embedding alignment evaluation checks consistency of vector spaces across model updates.

When embedding models are updated, vector spaces may rotate or shift, making previous embeddings incompatible. Alignment evaluation ensures semantic consistency using techniques like Procrustes analysis or anchor-based mapping. Without alignment, downstream systems like search or clustering degrade significantly.