What do you mean by span compression? We have experimented with various embedding contes lengths and we have found that bigger embeddings aren't the ones provid

What do you mean by span compression? We have experimented with various embedding contes lengths and we have found that bigger embeddings aren't the ones providing the best recall. We have hit the best results with something between 65% to 75% of the maximum embedding context length. We have been using OpenAI embeddings models though.