PDF to JSON Converter — Free & Online

What can you do with PDF as JSON?

Machine-readable structure instead of a wall of text.

How to convert PDF to JSON

Three steps and a few seconds.

1. 1

   Select your PDF

   Drag and drop the file or browse your device. Any size, any page count.

2. 2

   Structure is extracted

   Each line is classified and tagged with its page, position and font size — on your own device.

3. 3

   Copy or download

   Copy the JSON or save it as a .json file, ready for your pipeline.

Need something simpler? Extract plain text for reading and indexing, or convert PDF to Markdown for structured prose.

The shape of the output

The result is a single object with fileName, pageCount, and a pages array. Each page holds an array of lines, and each line carries four things: a type of heading, list_item or text; the text itself; the fontSize it was drawn at; and its vertical position on the page.

Line types are inferred rather than declared, because a PDF does not record them. A line set noticeably larger than the surrounding body text is classified as a heading; a line beginning with a bullet or numeric marker is classified as a list item; everything else is text. The font size is included precisely so you can second-guess that classification and apply your own threshold.

The vertical position is what makes reconstruction possible. Large gaps between consecutive lines usually mean a section break, and lines sharing a position across many pages are almost certainly running headers or footers — both of which are much easier to detect with coordinates than with text alone.

Chunking PDFs for RAG

The default approach to retrieval-augmented generation — split the text every N characters with some overlap — is fast and mediocre. It cuts sentences in half, splits tables from their captions, and produces chunks that begin mid-thought, which is exactly the sort of context an embedding model handles worst.

Typed lines let you do better. Start a new chunk at every heading, and each chunk becomes a section that stands on its own. Carry the heading text into the chunk so the embedding captures what the section is about, not only what it says. Merge short chunks with their neighbours, and split overlong ones at the nearest line break rather than mid-word.

Because the page number travels with every line, citations come free: when the model retrieves a chunk you can tell the user which page of which document it came from. That is usually the difference between an answer a reader trusts and one they cannot check.

Why this runs in the browser and not behind an API

The commercial document-parsing services all work the same way: you upload the file, they parse it, you pay per page. That is a reasonable trade for public documents and an impossible one for the documents that most need parsing — contracts under NDA, patient records, unreleased financials, anything where "we sent it to a third party for processing" is the sentence you cannot write.

Running pdf.js in the browser removes the trade entirely. There is no key to rotate, no per-page bill, no rate limit, no data-processing agreement to negotiate, and no vendor retention policy to read. Because there is no server round trip, the tools keep working with the network disconnected. Load the page, turn off Wi-Fi, and the conversion still runs — which is the most direct proof available that nothing is being uploaded.

Frequently asked questions