Can someone explain the embeddings module for GPTIndex? How is it different from doing cosine similarity across a set of vectors and grabbing top k results and injecting that into the prompt?