Universal Document Parsing

Stop parsing
photos of your
documents.

Your DOCX is already structured XML. AILANG reads it directly — full structure preserved, no AI calls, no per-page bill.

Free, in your browser — no account required

PDF-first parsers
DOCX structured XML
→
PDF
track changesmerged cellscomments
→
ML rebuild 2–5 sec
→
Flat text
OfficeDocBench v2 comparison
AILANG Parse
DOCX structured XML
→
Direct XML instant
→
Typed blocks
track changesmerged cellscomments
→
Your format
JSONMarkdownHTMLQuarto+5
98.1% OfficeDocBench — how we test
The Problem

PDF-first parsers destroy what's already there

Converting DOCX to PDF is like photographing a spreadsheet. Track changes, merged cells, comments — all gone. Then ML spends seconds reconstructing what the XML already had.

HeadingBlock
TextBlock
TableBlock
ChangeBlock
ImageBlock
SectionBlock
Lost in PDF conversion

Track changes: kept, not flattened

See who changed what and when. Author, timestamp, original vs revised text. The audit trail that vanishes when you convert to PDF.

On OfficeDocBench, AILANG Parse returns tracked changes in 3 of 3 files that contain them. Docling, Unstructured, LlamaParse and MarkItDown return them in none.

Payment terms: net-30 net-60 days from invoice date.Alice Chen · Mar 15
Delivery deadline: Q2 2026 Q3 2026Bob Martinez · Mar 16
Liability cap: $500,000 $1,000,000Alice Chen · Mar 17
Lost in PDF conversion

Merged cells: kept, not flattened

You get the exact colspan and rowspan of every merged cell. Parsers that go through PDF split merged cells into separate cells, which breaks the table.

PDF-first: flattened
RegionQ1Q2
EMEAUK$125K$140K
DE$98K$112K
→
AILANG Parse: kept
Region colspan:2Q1Q2
EMEA rowspan:2UK$125K$140K
DE$98K$112K
OfficeDocBench

87 files. 11 formats. Open-source.

11 parsers, the same files, the same scoring. A file a parser fails on or doesn’t support counts as zero.

AILANG Parse98.1%
Kreuzberg64.6%
Docling62.2%
Reducto (r-1)60.1%
LlamaParse (standard)59.0%
Extend (Parse 2.0)53.3%
Unstructured51.4%
Raw OOXML48.7%
MarkItDown46.6%
Pandoc35.6%
LiteParse29.0%

Full methodology and results → · Word files side by side with MarkItDown →

In your AI assistant

Use it inside ChatGPT, Claude and Codex

Add one URL and sign in with Google or GitHub. Your assistant can then read, convert and edit Word, Excel, PowerPoint and PDF files with their structure intact. No API key to copy, no code to write.

How we build it

Benchmarks first,
code second

Most parsers are written first and measured later. We write tests with hand-checked answers first, then AI writes code to pass them. The high scores are the goal of that process, not a coincidence.

How we build against the tests →

01
Write the tests
28 structural tests
hand-checked answers
→
02
AI writes parser code
AILANG code
written to pass each test
→
03
Run benchmarks
tables, track changes
merged cells, comments
=
98.1%
Composite score
Privacy by Default

Office files stay in your browser

DOCX, PPTX, XLSX and other Office formats are parsed in your browser, or on your own machine with the CLI. No upload, no server, nothing sent.

PDFs are parsed locally by default. Images need AI: your key, your provider, your choice. The hosted API processes files on our servers and keeps only short-lived copies, which are deleted automatically within about a day. Privacy policy →

Try it now ↓

100% Local Browser & CLI
.docx .pptx .xlsx
→
WASM / CLI
→
Typed blocks
Nothing leaves your machine. No server, no upload, no AI needed.
Encrypted API — all formats
.pdf .docx .png
→
API Server
→
Typed blocks
Processed on our servers; short-lived copies are deleted automatically. Office formats don't use AI, even through the API.
Universal

17 formats in, one structure out

Every format comes back as the same typed blocks. Office files are parsed without AI. PDFs are read without AI by default, with an AI backend for scans. Images use an AI model: Gemini, Claude, OpenAI, or Ollama running locally.

.docx
.pptx
.xlsx
.odt
.html
.md
.csv
.tex
.pdf
.png
AILANG Parse
TextBlock
HeadingBlock
TableBlock
ImageBlock
ChangeBlock
98.1%
OfficeDocBench
~100×
Cheaper
14
Input formats
1
Dependency
Try It Now in your browser
Loading WASM runtime...
All parsing happens locally via WebAssembly. Your files never leave the browser.
Try a sample: See a 421-page SEC filing, parsed → Word to Markdown → PDF back to Word → PDF back to PowerPoint →

Drop a file here

Or click to browse

Local: DOCXPPTXXLSXODTODPODS HTMLMDCSVEPUBEMLTEXRTF
AI key: PDF*PNG*JPG* MP4*MP3*WAV*

AI Settings

Unlocks AI parsing of image files and scanned PDFs. Pictures inside documents are kept as images.

Stored only in your browser, never sent to our servers. Get a free key at aistudio.google.com.
Original Parsed JSON Markdown A2UI
Drop a document or try a sample to see it parsed
Need to parse more than one file?
Open the Workbench for a multi-file library, saved insights, and CLI and SDK snippets to copy. Same parser, more room.
Open Workbench →

1
Try it
Parse documents in your browser. No signup, no install.
Open Workbench
2
Use the API
1,000 requests/month free. Python, JS, Go SDKs.
Get API Key
3
Install locally
Unlimited local parsing. Only requires AILANG.
Install CLI
Free
€0
1,000 req/mo · 50 AI parses
Get API Key
Pro
€29/mo
100,000 req/mo · 500 AI parses
Start Pro
Business
€99/mo
500,000 req/mo · 2,000 AI parses
Start Business
Custom
Let’s talk
Higher volumes · on-prem · SLAs
Get in touch
Per document, not per page
Each request parses a whole document: a 200-page DOCX costs the same as a 1-page DOCX.
Pro is ~100× cheaper: per-page APIs charge $25–$250 for 1,000 docs of 25 pages each. No per-page fees. See full pricing →
Tip: Send the original Office file (DOCX, PPTX, XLSX) instead of converting it to PDF. Office parsing is exact, instant, and doesn’t count against your AI quota. A 5 MB DOCX typically becomes a 10–30 MB PDF once fonts are embedded. That costs more, takes longer, and uses AI when none is needed. Save your AI parses for scanned documents and images. Only have the PDF? Turn it back into Word or PowerPoint in your own template.

The CLI and the in-browser parser are free and unlimited, including AI parsing with your own key. Pricing applies to the hosted API only. Full pricing details →


What Each Parser Keeps
What is keptAILANG ParseDoclingUnstructuredLlamaParseMarkItDown
Tracked changes returned3 of 3 files0 of 3 files0 of 3 files0 of 3 files0 of 3 files
Comments returned2 of 2 files0 of 2 files0 of 2 files0 of 2 files0 of 2 files
Headers and footers returned3 of 3 files0 of 3 files1 of 2 files0 of 3 files0 of 2 files
Text boxes returned2 of 2 files0 of 2 files0 of 2 files0 of 2 files0 of 2 files
Speaker notes returned1 of 1 files0 of 1 files0 of 1 files1 of 1 files0 of 1 files
Table cells: position, merges, formulas100.0%81.6%78.1%78.2%54.5%
List type and depth80.0%80.0%0.0%45.7%29.2%
Overall score, on the files it could read98.1%65.2%58.9%61.9%58.7%
Overall score, counting unreadable files as zero98.1%62.2%51.4%59.0%46.6%

Measured on OfficeDocBench: in how many of the test files containing each feature the parser returned it. Read from the published results.
Full benchmark methodology and results →


Supported Formats 17 formats

Office formats are read directly from their XML. Text formats such as Markdown, CSV and HTML are parsed directly too. For PDFs, images, audio and video, use any AI provider: Gemini, Claude, OpenAI, Ollama (fully local), even LlamaParse or Unstructured. You get the same structured output whichever provider you choose. No vendor lock-in.

Deterministic (read from the XML, no AI) Native (parsed as text, no AI) AI-powered (any provider) Generate only (output format)

Structural Extraction what reading the XML gives you

Many parsers convert Office files to PDF first and lose structure. AILANG Parse reads the XML directly, and these features depend on that.

Track Changes
Insertions, deletions and moves, each with its author and timestamp. You get the changes as data, not just the accepted or rejected text.
From the XML
Merged Cells
Table headers, rows and cells keep their colspan and rowspan. Merged cells are not split into separate cells.
From the XML
Headers & Footers
Returned as their own sections, separate from the body text.
Better
Text Boxes & Shapes
Text inside shapes and text boxes (DrawingML and VML), including older VML images that other parsers often drop.
From the XML
Embedded Images
Images returned with their data and type. Optional AI descriptions from Gemini, Claude or Ollama.
Better
Verified Contracts
A guard against silent data loss. 63 contracts are verified when the parser is built. For example, a step that converts a list of blocks cannot drop or add one.
Unique

What You Get real output from real documents

Every format comes back as the same typed blocks. Track changes, merged cells, comments and more are kept as data, not flattened into text.

track_changes_move.docx From the XML Moved paragraphs kept, with author & timestamp
Author: Jesse Rosenthal Created: 2016-04-16 Changes: 2 tracked

Here is some text.

Here is the text to be moved.

↘ Moved hereJesse Rosenthal · 2016-04-16
Here is the text to be moved.

Here is some more text.

↖ Moved from hereJesse Rosenthal · 2016-04-16
Here is the text to be moved.
View JSON
{
  "type": "change",
  "changeType": "move-to",
  "author": "Jesse Rosenthal",
  "date": "2016-04-16T08:20:00Z",
  "text": "Here is the text to be moved."
}
merged_cells.docx From the XML Column spans & merged cells kept
Author: Shay Hill Created: 2023-01-23
0-0 0-12 colspan: 2 0-3
12-01-11-21-3
merged ↑2-12-22-3
3-034-123 colspan: 3
4-0merged ↑ · colspan: 3
View JSON
{
  "type": "table",
  "headers": ["0-0", { "text": "0-12", "colSpan": 2 }, "0-3"],
  "rows": [
    ["12-0", "1-1", "1-2", "1-3"],
    [{ "colSpan": 1, "merged": true }, "2-1", "2-2", "2-3"],
    ["3-0", { "text": "34-123", "colSpan": 3 }]
  ]
}
tables.docx Structural Headers, rows, and multi-paragraph cells

A table, with and without a header row

NameGameFameBlame
Lebron JamesBasketballVery HighLeaving Cleveland
Ryan BraunBaseballModerateSteroids
Russell WilsonFootballHighTacky uniform
SinpleTable
WithoutHeader
Simple

Multiparagraph
Table

Full
Of

Paragraphs
In each

Cell.
View JSON
{
  "type": "table",
  "headers": ["Name", "Game", "Fame", "Blame"],
  "rows": [
    ["Lebron James", "Basketball", "Very High", "Leaving Cleveland"],
    ["Ryan Braun", "Baseball", "Moderate", "Steroids"],
    ["Russell Wilson", "Football", "High", "Tacky uniform"]
  ]
}
comments.docx From the XML Comments kept, with their authors
Author: Jesse Rosenthal Created: 2016-05-09 Comments: 5

I want some text to have a comment on it.

This is a new paragraph.

And so is this.

One more. And this is one with a comment in a comment.

JR
Jesse Rosenthal
I left a comment.
JR
Jesse Rosenthal
A comment across paragraphs.
JR
Jesse Rosenthal
This one has multiple paragraphs. See?
JR
Jesse Rosenthal
Do something.
JR
Jesse Rosenthal
Do something else.
View JSON
{
  "type": "section",
  "kind": "comment",
  "blocks": [
    { "text": "[Jesse Rosenthal] I left a comment." }
  ]
}
interview.mp3 AI-powered Transcription with speaker detection & language ID
interview.mp3
audio/mp3
2 speakers Language: en 4 blocks extracted
“Welcome to the show. Today we're discussing document parsing and why structure matters more than text extraction...”
Summary: Discussion about document parsing approaches
View JSON
{
  "type": "audio",
  "transcription": "Welcome to the show. Today we're discussing...",
  "mime": "audio/mp3"
}
tutorial.mp4 AI-powered What is shown and said, returned as structured blocks
tutorial.mp4
video/mp4 · 4 blocks extracted
Technical tutorial showing a spreadsheet with quarterly sales data, followed by a presenter explaining trends

Q1 Sales Review

RegionRevenueGrowth
North America$2.4M+12%
EMEA$1.8M+8%
“As you can see from the data, North America led growth this quarter...”
View JSON
{
  "type": "video",
  "description": "Technical tutorial showing a spreadsheet...",
  "mime": "video/mp4"
}
Real output from actual test documents. Track changes, merged cells and comments are kept as structure, not flattened to text. Expand “View JSON” on any tab to see the raw blocks.

Benchmarks OfficeDocBench
98.1%
AILANG Parse composite
100.0%
Format coverage
64.6%
Kreuzberg (adjusted)
48.7%
Raw OOXML (adjusted)
35.6%
Pandoc (adjusted)
Metric AILANG Parse 0.58.0 Raw OOXML Pandoc 3.12.1 Kreuzberg 4.10.4 MarkItDown 0.1.5
Composite score 98.1% 70.7% 63.1% 68.5% 58.7%
Coverage-adjusted 98.1% 48.7% 35.6% 64.6% 46.6%
Format coverage 100.0% 69.0% 56.3% 94.2% 79.3%
Track changes (files returned / files containing it) 3/32/33/30/30/3
Comments (files returned / files containing it) 2/22/20/20/20/2
Headers & footers (files returned / files containing it) 3/32/20/32/20/2
Text boxes (files returned / files containing it) 2/21/20/20/20/2
Equations (files returned / files containing it) 1/10/10/10/10/1
Runtime dependencies AILANG only Python stdlib Pandoc binary Python + libraries Python + libraries
Open source and reproducible
We build AILANG Parse against these tests, so a high score is the goal, not a coincidence. 87 test files in 11 formats, taken from Pandoc, Apache POI, LibreOffice, Unstructured, python-pptx, and Project Gutenberg rather than written to suit us. The expected answers, the scoring code and the wrapper for every parser are published.
Most recent run: 2026-11-18. Adjusted score = composite × (files parsed / total files). AILANG Parse is the only parser tested that reads every file.
View on GitHub  ·  How we build against the tests
Found a document we don't parse well? Send it to docparse@sunholo.com. If we add it to our test set, you get 1 month of Business tier free (€99 value). Details →

How It Works & CLI Usage

Every format goes through the same steps. AILANG Parse detects the file type, picks the right parser, and returns the same typed blocks.

Document
.docx .pdf .epub ...
→
Detect format
By file extension
→
Parser
16 AILANG modules
→
Typed blocks
9 block types
→
Output
JSON + Markdown
TextBlock HeadingBlock TableBlock ImageBlock ListBlock SectionBlock ChangeBlock AudioBlock VideoBlock

Usage

One command. Every format. CLI or library.

Office Documents (no AI needed)
# Parse any document (Office formats — instant, no AI) docparse report.docx docparse slides.pptx docparse data.xlsx docparse document.odt docparse book.epub
PDFs (no AI by default) and images
# PDF: deterministic pdftotext by default — no AI, no API key docparse invoice.pdf docparse scan.png # images auto-select AI # Scanned/image-only PDFs escalate to local docling by default; AI is opt-in docparse scan.pdf --pdf-backend ai --ai gemini-3.5-flash-lite # Google docparse scan.pdf --pdf-backend ai --ai claude-haiku-4-5 # Anthropic docparse doc.pdf --pdf-backend docling # local ML structure
Dev Tools
docparse --check # Type-check all 18 modules docparse --test # Run 51 inline tests docparse --prove # Z3 contract verification docparse report.docx --verify # Runtime contract checks docparse report.docx --describe # AI image descriptions docparse report.docx --summarize # AI document summary

Parse a document.
Keep the structure.

Free in your browser, no account required. 1,000 API requests/month free. CLI and WASM are unlimited.

Try AILANG Parse →