Presumably, they're referring to chunking up the data into discrete semantic units—smaller vectorizable subsections (e.g., paragraphs) more precisely capturing different parts of the data.