Hello,

During the first iterations yes, the item and candidate representations will be probably far away. But the purpose of the optimization procedure is to make them close when it is relevant and far away when it is not. Because of this and because they both live in a represenation space of the same dimension, we can consider that they live in the same representation space