Document to Markdown — PDF, Word, PowerPoint and Excel for RAG

Converts documents into clean Markdown that a retrieval pipeline can chunk and embed, using Docling, IBM Research's open-source document converter. Reading order, headings, lists and table structure are recovered from the page layout rather than guessed from the text stream, which is the difference between a table that survives chunking and one that arrives as a column of stray numbers.

Give it a list of URLs. Get back one record per document with the converted text and the page count it was billed on.

Input

Field	Type	Default	What it does
documentUrls	array	—	Direct links to the files. PDF, DOCX, PPTX, XLSX, HTML, Markdown, AsciiDoc, CSV and images (PNG, JPEG, TIFF, BMP, WebP). Up to 200 per run. A URL with no extension is accepted — Docling reads the bytes.
outputFormat	select	markdown	markdown, html, json, or text. JSON keeps Docling's full layout tree, including reading order and table cell structure.
ocr	boolean	false	Read scanned pages with OCR. Leave off for digital PDFs: OCR is several times slower and reads worse than the text layer that is already in the file.
maxPages	integer	—	Stop each document after this many pages. Billing counts only the pages actually converted.

Output

One record per document:

source_url        the URL it was fetched from
filename          last path segment, query string stripped
doc_format        pdf, docx, pptx, xlsx, html, md, asciidoc, csv, image
page_count        pages converted — null for formats that have no pages
character_count   length of `content`
content           the converted document, in `output_format`
output_format     which format `content` is in
ocr_used          whether the text was read off the page or lifted from a text layer
tables            tables Docling recovered
pictures          figures Docling recovered
title             the document's own title, when the file carries one

Two of those are worth reading carefully.

page_count is null, never 0, for HTML, Markdown, AsciiDoc and CSV. Those formats have no pages at all, and a zero there would read as a measurement — an empty document — rather than as an inapplicable field. The run still bills one unit for them, because the conversion still costs compute.

ocr_used tells you which of two very different error profiles the text has. A digital PDF's text layer is exact. OCR output is a good guess, and a RAG index built on the second while assuming the first will return confident answers from misread numbers.

Billing

Pay per event, plus a small actor-start fee covering the cold start and model load.

Event	When
page-converted	One page read from the document's own text layer.
ocr-page-converted	One page read with OCR. Priced higher because it is measurably more work — on the same 15-page file, OCR cost about three times as much to run.

The event follows what actually happened, not what you asked for: a document Docling converted from its text layer is charged the cheaper event even on a run with OCR switched on.

• A document that fails to convert is never charged. Nothing is billed until the record is in the dataset.
• A URL whose extension Docling cannot read is rejected before anything is downloaded, so it costs nothing.
• The same URL listed twice is converted once and billed once.
• Pageless formats bill one unit each.

What it does not do

• It does not crawl. Give it the file URLs; it does not follow links to find them.
• It does not read files behind a login. A signed download URL works; a page that needs a session does not.
• It does not translate, summarise or otherwise change the text. The output is the document, restructured.
• It does not fabricate a page count for a format that has none. See page_count above.

Notes on accuracy

Docling recovers layout with a page-layout model and a table-structure model (TableFormer), so multi-column pages, headers and footers, and merged table cells come through in reading order. What it cannot do is invent information the page does not carry: a scanned page with no OCR yields no text, and a table drawn as an image is a picture.