Built on IBM Granite Docling

Convert PDFs and images to clean Markdown, JSON, and HTML. Nothing uploads.

Parsade runs IBM's open-source Docling pipeline locally in your browser via WebGPU and Pyodide. Your documents never leave your device.

Free · no account Works on PDFs, images, and scanned documents.

No sign-up, no page limits. First load fetches ~1.15 GB of model weights (less for smaller builds), cached after that.

The pipeline · everything in this browser tab
01 · input
pdf.js
Rasterizes PDF pages to canvas bitmaps. Images pass through.
02 · vision
Granite Docling 258M
Runs on WebGPU (~1.15 GB), emits DocTags layout markup.
03 · convert
docling-core
Python in Pyodide (~25 MB) builds a structured DoclingDocument.
04 · output
Structured files
.md .json .html .doctags RAG chunks

Transparent by design.

On privacy, on provenance, on the model that runs. Look through the glass.

01

Privacy-first by architecture

The free tier has no backend. Documents are parsed on your GPU and processed by Python in WebAssembly, all in this tab. Lawyers, analysts, healthcare engineers: open dev tools and verify the network panel stays empty.

02

The official Docling library

The browser runs IBM's official docling-core Python library via Pyodide. Markdown, HTML, and structured JSON are byte-identical to what the canonical CLI produces server-side.

03

Batch, queue, ZIP

Drop a folder's worth of PDFs and images at once. Parsade queues them, processes them sequentially on the GPU, and lets you download every output as a single ZIP archive.

04

Hardware-aware

A built-in WebGPU check measures your GPU's actual throughput and tells you whether conversion will be fast, moderate, or slow, flags the software fallback that means no GPU acceleration, and shows what to do about it.

How it compares

One of a handful of tools built for PDF-to-RAG. Privacy, license, provenance, and chunking side by side.

Questions

Is my data really private? +

Yes. There is no backend. Models load from a CDN once, then everything runs locally. Verify it yourself: open DevTools → Network, drop a document, and watch nothing leave your device.

Why does the first load take so long? +

On first visit Parsade downloads the model weights (about 1.15 GB for the default Full quality, down to ~262 MB if you pick a smaller build) plus roughly 25 MB of Python runtime. Both are cached in IndexedDB and load instantly after.

Which browsers work? +

Chrome, Edge, and Brave 113+ with WebGPU enabled. Safari 26+ (macOS Tahoe / iOS 26) also works. Firefox WebGPU is still maturing; results vary.

How is this different from docling-serve? +

docling-serve is IBM's reference Python service that you self-host. Parsade packages the same conversion stack into a static webpage that needs no server at all. Same output, no server to run.

Why Pyodide instead of a JS port? +

IBM ships the Markdown/HTML conversion only in Python (docling-core). The official JS package, @docling/docling-core, is just TypeScript types. By running real Python in WebAssembly, Parsade gets byte-identical output to the canonical library without re-implementing the logic.

Can I use this offline? +

Once everything is cached (after the first load), yes. Open the tab with no internet and conversion still works.

Convert a document now.

Free, no account. Drop a PDF or image and get clean Markdown, JSON, and HTML in your browser. First load fetches ~1.15 GB of model weights, cached after that.