How would you answer an interview scenario involving Embeddings?

For a knowledge base with millions of chunks, describe how you would migrate to a new embedding model safely and verify retrieval quality.