Ida

Pre versus post filtering is barely advertised and it decides everything above a few thousand documents.

Changed embedding model twice and the second time was painless purely because of how the first one went.

What makes it survivable: store the raw text and the metadata as the source of truth, treat the vector store as a derived index you can throw away, and keep the model name on every row. Then a migration is build the new index alongside, switch, delete the old one.

If your chunks only exist inside the vector store, a model change is a data recovery exercise.