I’ve done some experiments and the ideal seems to be semantic chunks of max 175 tokens. But only if those chunks are already very dense (use summarisation to get there)