In RAG document preprocessing, text chunks are typically split into 200–1,000 tokens; chunks that are too large introduce noise, while chunks that are too small lose context.