It's not obvious to me that vector size isn't relevant in semantic search. What is it about the training process for semantic search that makes that the case?