How much compute do your systems expend on chunking vs. the embedding itself?