embedding model migration: old vectors, new queries, no matches

I am seeing teams swap embedding models and reindex in place. The new model scores better on the public retrieval benchmark. The rollout still breaks retrieval for days. Each embedding model draws its own private map of meaning. A query embedded by the new model lands on the new map. Half the index still sits on the old map. Similarity across two maps is noise that looks like a score. Nothing throws an error, and the top results just get quietly worse. The fix is a second index, not a mixed one. Build the new index beside the old one. Send each request to exactly one index, never to both. Switch over only when the backfill reaches every document. Tag each vector with its model version and refuse mixed reads. The trade-off is double storage and double write cost during the move. Measure recall on a fixed query set against both indexes. Also track the share of documents still on the old version. That share should reach zero before the switch, not after.