Use semantic chunking

Try to split the documents not into chunks in some static manner, but in a way that leaves the already existing semantic chunks in a document (paragraphs) mostly intact.