How do you change the embeddings model without breaking search?
Answer out loud or in your head first, then compare.
Answer
I create a new version of the index and compute the documents' representations with the new model. Queries must be encoded with the same new setup. I don't mix vectors just because they have the same size.
Before switching over, I check the data is complete and the quality on the same questions. I can compare both variants for a while, keeping the option to go back. I also have to account for documents added or deleted while the new version was being built. If I change the model, chunking and ranking at the same time, it's harder to find out what helped or hurt. So I try to check the most important variables separately.
Why it matters: An example of a real change in a running system, once embedding basics are mastered.
Possible follow-ups: What about document updates during a long index rebuild?
How did it go? Be honest — “Not yet” puts the question on your review list.
Start with
- G-P01 · What is RAG? — RAG and retrieval
- G-P02 · What are embeddings? — RAG and retrieval
Further reading
- Claude Platform: Embeddings(opens in a new tab)Docs; accessed 22.10.2026. Semantic representations and how to choose an embedding model.
- pgvector: official repository and documentation(opens in a new tab)Docs; accessed 22.10.2026. Exact vs ANN search, indexes, and how filters affect search.