Every element, every page
Role, box, reading order, confidence and typed content: titles, text, tables, formulas, figures, captions, footnotes.
Solutions · Document Intelligence · Document parsing
Parse any document. Keep every structure.
State-of-the-art AI parsing for .NET: PDFs, scans and images become typed, positioned elements, with charts read as data. Fully private.
report.json
{
"category": "figure",
"reading_index": 4,
"confidence": 0.97,
"bbox": [53, 70, 547, 329],
"content": {
"type": "chart",
"caption": "Revenue by region",
"text": "| Region | 2024 | 2025 |\n|---|---|---|\n| North | 12.4 | 15.1 |\n| South | 9.8 | 11.2 |"
}
}
A text layer is not a document.
Flat extraction returns words. The answers your systems need live in the structure around them, and that is what the parser keeps.
Charts
Read as data
Every bar, line and pie chart comes back as the table it plots, with its title, legend and panels.
Tables
Merged cells kept
Row and column spans survive as HTML, so a financial statement keeps its columns and its totals.
Order
Reading order
Columns, sidebars, captions and footnotes read in the order a person reads them, never interleaved.
Position
A box for every element
Bounds and a confidence on each element make citations, highlighting and review possible.
One setting. The engine does the rest.
Choose an effort level. The parser selects its models, verifies each one, and improves from one release to the next without a code change.
- Low. The least time per page, for high-volume ingestion where throughput leads.
- Medium. The balance of time and fidelity that suits most documents; the default.
- High. The most faithful read, for documents where every table, chart and heading counts.
- Verified models. Only the models the parser is tested with run; any other is refused before a page is read.
- Offline ready. Download the models once, or ship the files with your application.
A few lines from file to data.
One parse renders every standard format, with figure pictures kept on request.
ParseAndExport.cs
using LMKit.Document.Parsing; using var parser = new DocumentParser(ParsingEffort.Medium) { IncludeFigureImages = true }; ParsedDocument document = parser.Parse("annual-report.pdf"); File.WriteAllText("report.md", document.ToMarkdown()); File.WriteAllText("report.json", document.ToJson()); document.SaveHtml("report.html"); document.SaveDocLangArchive("report.dclx");
Every chart of a document, in reading order, as the data table it plots.
ChartsAsData.cs
using LMKit.Document.Parsing; using var parser = new DocumentParser(ParsingEffort.High); ParsedDocument document = parser.Parse("filing.pdf"); foreach (ParsedPage page in document.Pages) foreach (ChartContent chart in page.Elements.Select(e => e.Content).OfType<ChartContent>()) { Console.WriteLine($"Page {page.PageNumber}: {chart.Caption}"); Console.WriteLine(chart.Text); // the plotted series as a table }
Store the lossless JSON and render any format later, without the source or the models.
ParseOnce.cs
using LMKit.Document.Parsing; // Parse once, store the result. File.WriteAllText("contract.json", parser.Parse("contract.pdf").ToJson()); // Later, on any machine: no parse, identical output. ParsedDocument stored = ParsedDocument.FromJson(File.ReadAllText("contract.json")); string html = stored.ToHtml();
Ship the model files with your application for air-gapped machines; each file is verified.
ShipTheModels.cs
using LMKit.Document.Parsing; // On a connected machine: fetch what an effort level uses. foreach (var card in DocumentParser.GetModelCards(ParsingEffort.High)) card.Download(); // On the air-gapped machine: point the parser at the files. using var parser = new DocumentParser(ParsingEffort.High, modelFiles);
Standard formats, not a proprietary result.
Every format renders from the same parse, so a heading, a table or a caption reads the same in all of them.
| Format | What it carries | Built for |
|---|---|---|
| Markdown | Headings, tables as HTML, formulas as LaTeX, charts as tables | LLM prompts and RAG chunks |
| JSON | Every element with category, box, reading order and confidence; a published schema | Pipelines and storage, in any language |
| HTML | Semantic markup with each element's category and box | Display, review and source highlighting |
| DocLang | The open AI-native document markup of the LF AI & Data Foundation | Exchange with DocLang tools and models |
Embed it, or serve it.
LM-Kit One serves the same parser at POST /lmkit/v1/document-parsing,
on the command line and as an MCP tool for agents.
Parsing feeds everything else.
Document to Markdown
The fast universal converter across PDF, Office, email and HTML, with a vision model of your choice.
Document RAG
Index parsed pages with their tables and chart data, and answer with page citations.
Structured extraction
Turn documents into typed fields with a confidence per field and a human-review flag.
Build it. Read it. Try it.
Working console demos on GitHub, step-by-step how-to guides on the docs site, and the API reference for the classes used on this page.
Demo
Structured Document Parser
Parse PDFs and scans into typed elements and export Markdown, JSON, HTML and DocLang.
Open on GitHub → DemoChart Data Harvester
Every chart of a folder of documents, read as data and written as CSV.
Open on GitHub → DemoParsed Document RAG
Questions over parsed documents, answered with page citations.
Open on GitHub → How-to guideParse Documents into Structured Elements
Effort levels, elements, exports, figure pictures, bring-your-own models.
Read the guide → How-to guideExtract Chart Data from Documents
Charts as data tables, panels and legends, exported as CSV.
Read the guide → API referenceDocumentParser
API reference for the AI document parser.
Open the reference →