Document Parsing
Turn PDFs, scans, slides and spreadsheets into clean text and tables a model can use, before chunking and embedding them.
Trends
Mistral AI gained the most ground in package downloads: 6.8× against the median npm package over a year.
Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.
- Oct 19 · Pricing Mistral AI changed its llm api pricing: Mistral Large 3 plan dropped; new plan Mistral Large 4.
The real choice
Hosted API or library you run. Hosted parsers handle messy scans, complex tables and odd layouts per page, with your documents going to their servers. Open-source libraries run locally for free and keep files private, but you provide the compute and accept lower accuracy on hard documents.
Pick by situation
Tap the ones that are you — the tools that fit light up below.- You only need the text out of born-digital PDFs, fast and free
pypdfpdf-parse
PyMuPDF - You want Markdown from many file types with one local tool
MarkItDown
Docling - Layout and tables matter and documents must stay on your own machines
DoclingMarker
- Your inputs mix PDFs, emails, HTML and office files and you may move to a hosted API later
Unstructured - Scans, dense tables or forms where a parsing error means a wrong answer
Reducto
LlamaParse
Mistral AI
Compare the tools
Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.
MarkItDownlibrary · open sourceConverting PDFs, Word, Excel, PowerPoint, HTML and more to Markdown with one small Python package.39 open source3.3M/wk
Use it whenYour documents are mostly born-digital and you want clean Markdown for an LLM with minimal setup.
Trade-offLight on layout analysis, so complex tables and scans come out worse than with model-based parsers.
No maker's product we track shows it for this yet · in 39 open-source projects.
Doclinglibrary · open sourceParsing PDFs and office files locally, with layout and table structure kept, at no cost.27 open source658.2k/wk
Use it whenDocuments are private or volume makes per-page pricing add up.
Trade-offRuns on your CPU or GPU, and heavy PDFs are slow without a GPU.
No maker's product we track shows it for this yet · in 27 open-source projects.
Unstructuredservice · open sourceMany file types through one pipeline, starting with the open-source library and moving to the hosted API later.1+47 open source743.8k/wk
Use it whenYour inputs mix PDFs, emails, HTML and office documents.
Trade-offThe open-source library is more basic than the hosted service, which bills per page.
Used by [LW24] Megaparse · in 47 open-source projects.
Shoutout for this awesome open-source project on top of which we built Megaparse.
Markerlibrary · open sourceAccurate PDF-to-Markdown with layout, tables, equations and OCR from local models, or the same through Datalab's hosted API.2 open source27.1k/wk
Use it whenYou want model-quality parsing on your own GPU, or a hosted API from the same code later.
Trade-offThe model weights are free only for research, personal use and startups under $5M in funding or revenue; slow without a GPU.
No maker's product we track shows it for this yet · in 2 open-source projects.
ReductoapiHard documents — scanned forms, dense tables, financial and legal files — where extraction accuracy matters most.

11+2 open source250.2k/wk4.1× vs npm
Use it whenParsing errors cause wrong answers your users will notice.
Trade-offPaid per page after starter credits, closed source, and documents leave your servers.
Used by August Law, Scale Ai Reducto, Slashy, Anterior, Benchmark and 6 more · in 2 open-source projects.
LlamaParseapiHosted parsing that uses layout models and LLMs on hard pages, with modes you trade off between cost and accuracy.17 open source61.9k/wk−56% vs npm
Use it whenYou build on LlamaIndex, or want a free monthly allowance before paying per page.
Trade-offCredit-based pricing where the accurate modes cost several times the basic one, and documents go to LlamaIndex's cloud.
No maker's product we track shows it for this yet · in 17 open-source projects.
Mistral AIapi · free tierOCR of scans and PDFs into Markdown with tables and images, from the same API as Mistral's text models.No evidence yet7.8M/wk6.8× vs npm
Use it whenYou already use Mistral, or want OCR priced per page without a separate document vendor.
Trade-offIt returns the page contents as OCR output, and pulling structured fields out is a second step with a model.
llms.txtIn ReplitNo maker's product we track shows it for this yet.
pypdflibrary · open sourcePulling text out of born-digital PDFs in pure Python with no native dependencies.142 open source28.3M/wk
Use it whenYou need basic text extraction in a Python app or serverless function and nothing more.
Trade-offNo OCR and little layout awareness, so tables and multi-column pages come out jumbled.
No maker's product we track shows it for this yet · in 142 open-source projects.
PyMuPDFlibrary · open sourceVery fast PDF text, image and table extraction in Python, with PyMuPDF4LLM for Markdown output.76 open source19.2M/wk
Use it whenYou process many born-digital PDFs and speed matters.
Trade-offAGPL-licensed, so closed-source products need a commercial license from Artifex.
No maker's product we track shows it for this yet · in 76 open-source projects.
pdf-parselibrary · open sourceExtracting text from PDFs in a Node.js app with one dependency.46 open source8.3M/wk+72% vs npm
Use it whenYour backend is JavaScript and your PDFs are born-digital.
Trade-offNo OCR, and layout-heavy pages lose their structure.
No maker's product we track shows it for this yet · in 46 open-source projects.
Before you choose
Look at the parsed output of your ugliest 20 documents before choosing — tables and multi-column layouts are where parsers differ. Keep page numbers and section headings in the output so answers can cite sources. Parse once and cache the result; re-parsing on every change is slow and costly.
- Using plain text extraction on scanned PDFs, which returns empty strings because the page is an image, not text.
- Flattening tables into running text, so the model can no longer tell which number belongs to which row and column.
- Checking parse quality on one clean sample and discovering the failures only after a user uploads a two-column scan.
Other options
Real choices most makers here won't need to weigh.
Decided alongside
What the 11 makers' products here chose for their other decisions.