Knowledge bases
Chunking methods
Thirteen ways to split a document — when to use which, and the formats each can read.
The chunking method decides the unit that gets searched and cited. Pick the wrong one and an answer will quote half a clause or a table row without its column names. The picker in the app shows each method's description and a Reads: … line; methods the deployment cannot run yet are greyed out with the reason.
| Method | For | How it cuts | Reads |
|---|---|---|---|
| General | Ordinary prose — the default | Splits at sentence boundaries and merges back to the token budget, with overlap. | pdf, docx, txt, md, html, csv, tsv, json, eml |
| Q&A | FAQs, support exports, interview transcripts | One chunk per question-and-answer pair; the question is embedded rather than the answer. | csv, tsv, md, txt, docx |
| Table | Tabular data | One chunk per row, each carrying its column names. Export XLSX to CSV/TSV first. | csv, tsv |
| One | Short documents where every part is context for every other — a one-page policy, a product sheet | The whole document as a single chunk. | as General |
| Paper | Academic papers | Splits on section names (Abstract, Method, Results…) and prefixes each chunk with its section. | pdf, docx, txt, md, html |
| Book | Books, long documents with chapters | Splits on chapters and numbered headings, prefixing each chunk with its heading path. | pdf, docx, txt, md, html |
Legal (laws) | Laws, decrees, contracts | One chunk per Điều / Article / 第N条, never merged with its neighbours. | pdf, docx, txt, md, html |
| Manual | Technical manuals, procedures | Splits on numbered sections and keeps the section number with the text, so a step still says which procedure it belongs to. | pdf, docx, txt, md, html |
| Presentation | Slide decks | One chunk per slide, never merged. Export PPTX to PDF first. | |
| Keeps From, To, Subject and Date with every chunk of the body. | eml, txt | ||
| Picture | Images, scanned pages | Needs a vision model — not available yet; convert to a text PDF. | — |
| Audio | Recordings | Needs speech-to-text — not available yet; upload the transcript. | — |
| Resume | CVs | Needs an entity-extraction model — not available yet; use General. | — |
Quick pick
- Not sure → General.
- The file is a list of questions and answers → Q&A.
- The file is a table → Table (exported to CSV).
- A law or a contract with articles → Legal.
- Answers must quote a short passage whole → One.
Check the result in the chunk list before bulk uploading.