In RAGFlow, what does choosing a chunk method per document actually change?

## The design choice behind chunk methods Most RAG stacks expose splitting as parameters: a chunk size, an overlap, maybe a separator list. RAGFlow exposes it as a **choice of template per document type**, and only then as parameters within that template. The reasoning is that the right cut points are a property of the document's genre, not of a number: a statute is cut at articles, a slide deck at slides, a spreadsheet at rows, a paper at sections. Picking the genre gets you most of the way; tuning the token count is the residual. ## The templates and what a chunk becomes - **General** — the default and the broadest in file-type support (office documents, PDF, text, Markdown, HTML, images). It uses DeepDoc's layout output, then combines pieces up to a configured token count, respecting delimiters. This is the closest thing to conventional splitting. - **Q&A** — the source already contains question-answer pairs; each pair becomes one chunk, with the question as the part that matches a user query. Built for FAQ and support-ticket corpora. - **Manual** — assumes a hierarchical, titled user manual and cuts at the section level, so a chunk is a coherent procedure rather than an arbitrary window. - **Table** — for spreadsheets and delimited files: the first row supplies the field names, and each subsequent row becomes a chunk carrying those field names with its values. - **Paper** — academic PDFs, split by section, which keeps abstract, method and results separable rather than smeared across arbitrary windows. - **Book** — long-form documents split at chapter/section granularity with a token cap so one chapter does not become one enormous chunk. - **Laws** — legal texts, cut at the article/clause headings that legal documents reliably carry. - **Presentation** — one chunk per slide, which matches how slide content is authored and read. - **One** — the entire document becomes a single chunk. Useful for short documents that only make sense whole, and for cases where you want the model to see the complete text. The exact set grows between RAGFlow releases, so check the dropdown in the version you are running rather than memorizing a list. ## Where the setting lives A chunk method is chosen when you create a knowledge base and becomes the default for everything uploaded into it. Each document in the file list also carries its own chunk-method column, so a knowledge base holding a mix of PDFs, slide decks and spreadsheets can apply the appropriate template per file. The chunk method also determines which configuration fields you are shown — token count and delimiters for General, and which of the enrichment toggles (auto-keyword, auto-question, RAPTOR) are applicable. ## Changing the method later This is the operational detail interviewers probe. Selecting a different chunk method does not retroactively re-cut anything: parsing is a job, and the new method only takes effect when you run parsing again for that document. That run replaces the document's chunks wholesale. Two consequences follow. First, any manual chunk edits, added keywords or disabled chunks for that document are lost — hand-fixing chunks and then re-parsing is wasted work, so fix the template first and edit only afterwards. Second, re-parsing a large scanned corpus is not free; DeepDoc runs again per page, so a template change on a big knowledge base is an hours-long job, which is a strong argument for sample-parsing a few representative documents before bulk upload. ## Choosing well The practical procedure: group your corpus by genre, not by folder. For each genre, upload two or three representative files, parse with the template that matches the genre, and read the resulting chunks. You are checking that chunks are self-contained (a chunk answers a question without its neighbours), that tables and captions are intact, and that no chunk is a fragment of two unrelated sections. If General with a token count is producing better chunks than the genre template, that is a legitimate finding — the genre templates assume the documents actually follow their genre's conventions, and a "paper" exported from a slide deck does not. ## The common misconception A chunk method is not merely a splitter setting. It changes what a chunk *is* — a row, a slide, a Q&A pair, an article, a section — and therefore what retrieval can return as a unit and what a citation points at. That is why RAGFlow keeps it a first-class per-document choice rather than a hidden parameter.