Models have a limit on how much text they can take in one pass, measured in tokens. Retrieval systems, which look up relevant passages to give a model context, work better with focused passages than with whole documents. Chunking prepares text for both. Size is usually set in tokens, and the right size depends on the use.
Where the cuts fall matters more than the exact size. A chunk that ends in the middle of a sentence or separates a question from its answer loses meaning. Documents are best split at headings and paragraphs, conversations between turns, and code between files or functions. Some pipelines repeat a little text across neighboring chunks so context is not lost at the boundary. Each chunk should keep the ID of its parent record and its sequence number, so the original can be rebuilt and results traced. For licensed data, ask the buyer before chunking. Many AI labs prefer whole records and apply their own chunking, and a dataset delivered only as chunks is hard to reassemble.
Record the chunking rule in the data card if chunks are delivered, including the target size and whether neighboring chunks overlap.