API · early access

PDF to Markdown and JSON API for RAG and LLM pipelines

Send a PDF, get back clean Markdown or JSON: text in reading order, paragraphs joined across pages, page numbers and running headers removed, tables as tables, recognized formulas as LaTeX. The same document model also produces HTML and EPUB. Everything runs on our own server, with no third-party AI service.

Planned price: $1 per 1,000 pages, the first 5,000 free. Strong on PDFs with a text layer; behind Marker on olmOCR-bench (57.3 against about 76), mostly on scans, old documents and formulas. Read the limits first.

Early access: keys are issued by hand

What you get back

The converter reads the position of every glyph in the PDF and rebuilds the document: columns, headings, paragraphs, lists, tables, formulas, footnotes and figures. Every output format is generated from that one model, so Markdown and JSON agree with each other.

Markdown

Headings, GitHub-flavored tables (HTML tables where cells are merged), inline and display math as LaTeX, footnotes as [^1], figures as image files with their captions. Optionally, a <!-- page 12 --> comment before the first block that starts on each new PDF page.

JSON

Metadata, a table of contents and one flat list of typed blocks in reading order: heading, paragraph, list, table (with header, colspan and rowspan), figure, math (LaTeX), code, footnote, note, quote. Each block has its page (counting from 1) and its bounding box on that page, [x0, top, x1, bottom] in points from the top-left corner, for citing and highlighting the source.

HTML

One semantic HTML file: <h1>-<h6>, <table>, <figure>, footnotes as asides.

EPUB and plain text

A reflowable ebook with a table of contents, or plain UTF-8 text with the same clean paragraphs.

Request and response

Start a job with POST /v1/convert, sending your key and the PDF. Conversion is asynchronous: you get a job id, poll its status, then download the result. The jobs run on the same queue as the converter on this site. The endpoint opens with the first keys, and the shape below may still change.

Planned request: send a PDF
curl https://wandpdf.com/v1/convert \
  -H "Authorization: Bearer $WANDPDF_KEY" \
  -F [email protected] -F format=markdown

{"id": "j_7f3c2a", "status": "queued"}
Planned request: poll the job
curl https://wandpdf.com/v1/jobs/j_7f3c2a \
  -H "Authorization: Bearer $WANDPDF_KEY"

{"id": "j_7f3c2a", "status": "done", "pages": 12,
 "result": "https://wandpdf.com/v1/jobs/j_7f3c2a/result"}

The output, on the other hand, is real. Below is unedited output for Deep Residual Learning for Image Recognition (He et al.), one of the papers in our test set. Markdown: a table with its caption.

resnet.md
|  | plain | ResNet |
| --- | --- | --- |
| 18 layers | 27.94 | 27.88 |
| 34 layers | 28.54 | **25.03** |

Table 2. Top-1 error (%, 10-crop testing) on ImageNet validation. Here the ResNets have no extra parameter compared to their plain counterparts. Fig. 4 shows the training procedures.

JSON: a heading, the paragraph after it and a footnote.

resnet.json
{"title": "resnet", "lang": "en", "pages": 12,
 "toc": [...],
 "blocks": [
  ...
  {"type": "heading", "chapter": 0, "level": 3, "id": "h6",
   "text": "3.2. Identity Mapping by Shortcuts",
   "page": 3, "bbox": [50.1, 562.6, 213.9, 572.5]},
  {"type": "paragraph", "chapter": 0,
   "text": "We adopt residual learning to every few stacked layers. A building block is shown in Fig. 2. Formally, in this paper we consider a building block defined as:",
   "page": 3, "bbox": [50.1, 581.4, 286.4, 614.3]},
  ...
  {"type": "footnote", "chapter": 0, "label": "2",
   "text": "This hypothesis, however, is still an open question. See [28].",
   "page": 3, "bbox": [61, 704.1, 257.8, 712.7]},
  ...
 ]}

Limits: where it falls short

We would rather you find out here than after a week of integration.

  • On olmOCR-bench we score 57.3 (November 3, 2026). Marker scores about 76 and olmOCR 82, as published by the benchmark's authors. Most of the gap comes from scans, old documents and formulas, where the leaders run vision-language models on GPUs and we do not.
  • Scans are the weakest part. Tesseract OCR reads clean scans of printed text, but handwriting, poor photocopies and formulas in scans come out badly or not at all.
  • Tables without ruling lines are detected by rules and a small model; Marker and Docling get more of them right.
  • Formulas in PDFs with a text layer usually come out as LaTeX, but some display formulas come out as plain text (both numbered formulas in the ResNet paper above do) or split into text and LaTeX pieces.
  • Where it is strong: PDFs with a text layer, such as books, reports, manuals and papers. On our test of 31 public PDFs, 25 WandPDF conversions had no major defects (split paragraphs, page numbers in the text, lost text), against 16 for Marker. We make WandPDF and wrote that test ourselves, so read its limits too.
  • Early access means no SLA. One server, no batch discounts yet, and a new API.

Price

Planned price: $1 per 1,000 pages, pay as you go, the first 5,000 pages free. Nothing is billed during early access, so you can try it on your own documents first. For comparison, list prices per 1,000 pages in November 2026:

ServiceModePer 1,000 pages
WandPDFone mode, CPU$1
Datalab (Marker)lite$0.50
Mathpixbatch files$1.50
Datalab (Marker)fast or balanced$4
MathpixPDF conversion$5
Datalab (Marker)accurate$10

Your files never leave our server

  • No OpenAI, Anthropic, Google or any other AI service sees your documents. The converter uses rules plus small layout and table models, running on the CPUs of our own server, and OCR for scans is Tesseract on the same machine.
  • The PDF is deleted as soon as it is converted, the result within an hour.

Get an API key

Tell us your volume and what you convert. We issue keys by hand and reply by email; the answer comes from a person, not an autoresponder.

We use your email only to reply about the API.