MarkPrep

Chunk preview for RAG

See how your document splits into chunks before you index it, with heading-aware chunking, size and overlap sliders, and JSONL export.

Drop files, paste, or click to browse
PDF, Word, Excel, PowerPoint, HTML, CSV, JSON, EPUB, images and more. Nothing is uploaded.
Live count of requests to other websites during conversion: 0To see for yourself, press F12, open the Network tab and convert a file.

Drop a document, open the Chunks tab and read every chunk as your retrieval pipeline will get it. Each chunk shows its heading path, its estimated token count and, for PDFs, its page number. Move the size slider (100 to 2,000 tokens) and the overlap slider (0 to 400 tokens), and the chunks redraw at once. The default is 800 tokens with 100 tokens of overlap. When the chunks look right, download them as a .jsonl file with one chunk per line.

Why chunking matters as much as the conversion: chunking is where RAG breaks.

Frequently asked questions

What chunk size should I use?+

Start with the default of 800 tokens and 100 tokens of overlap. A block such as a table is split only when it alone is bigger than the chunk size. If a long table is cut in two, raise the size. Heading-aware chunking suits Markdown, because Markdown marks each section with a heading.

What's in the JSONL?+

Every line of the .jsonl file holds one chunk as a JSON object with three keys: id, text and metadata. The metadata holds the source file name and title, plus chunk_index, heading_path, page and tokens. The page field is the page number for PDFs, the slide number for PowerPoint, and null for other files. For EPUB it is the number of the file the text came from, because an EPUB stores a book as several files.

How does heading-aware chunking work?+

Once a chunk is at least 80% full, the next level 1 or level 2 heading (# or ##) starts a new chunk. That keeps each of those headings with the section under it. You can switch heading-aware chunking off.

Why don't chunk sizes change when I pick another model?+

Chunk sizes use one model-neutral estimate of about 3.5 characters per token, so the boundaries stay the same when you pick another model.

Does my document upload?+

No. Conversion and chunking run in your browser, and the JSONL download is created on your device.