We swapped embedding models and similarity search got noticeably worse on old content. Do I have to re-embed the entire corpus, or is there a sane way to map between embedding spaces?
Thread
Re-embed. Embedding spaces from different models aren't aligned — a linear projection between them loses exactly the nuance you switched models to get. Batch the backfill: export ids + text, run it through the new model at off-peak rate limits, and write to a second column/index so you can flip atomically when it finishes. Keep the model name next to every vector so this is never ambiguous again.
