I have a document search feature built with OpenAI embeddings, and I’m considering switching embedding models. The existing index contains several hundred thousand chunks, so regenerating everything would take some planning.

What I’m unsure about is compatibility. If the new model can produce vectors with the same dimensions as the old one, does that make them comparable, or could they still represent meaning in completely different ways?

What happens during the changeover?

• Can new documents use the new model while older documents keep their existing vectors?
• Would queries need to run against two separate indexes during the migration?
• How should I compare retrieval quality without confusing model differences with changes to chunking?

I’d appreciate advice on embedding model migration from anyone who has done this on a live system. My main concern is avoiding a period where search still returns plausible results but quietly misses the most relevant documents.