Columns
Two arguments, interleaved
Read line by line across the page, two columns merge into sentences no author wrote.
Use case
Papers your LLM can read. Formulas, figures and all.
Two-column papers, equations and plots become structured Markdown and JSON: formulas as LaTeX, plots as data, every claim traced to its page.
A paper is not a wall of text.
Literature search and research assistants fail on the same pages: two columns, equations and figures that text extraction turns into noise.
Formulas
Equations as symbol soup
Subscripts, Greek letters and fractions come out flattened, and the result loses its meaning.
Figures
Results trapped in plots
The key result is often a curve or a bar chart, invisible to anything that reads only text.
The structure of the paper, kept.
The parser returns the paper as typed elements in reading order, ready for Markdown, JSON, HTML or DocLang.
Order
Read like a person
Columns, captions and sidebars in the order the author meant, footnotes and running heads kept apart.
LaTeX
Formulas you can render
Display equations come back as LaTeX, ready for a renderer, a notebook or an LLM prompt.
Data
Plots as tables
Bar charts, line plots and multi-panel figures return the values they plot, with legend and units.
Tables
Results tables, whole
Grouped headers and merged cells survive as HTML spans, so each value stays in its column.
Outline
Sections across pages
Heading levels are ranked across the whole paper, so a section keeps its level on every page.
Grounding
Every claim has a source
Each element carries its page and box, so an answer can cite the exact paragraph, table or figure.
Ask the literature. Get cited answers.
Parse once, index the elements, and a research assistant answers from the papers themselves, pointing to the page and region behind each claim.
- Chunks that follow the paper. Split on sections; a table or a figure stays whole.
- Values you can retrieve. Plot data and table cells become text an index matches.
- Answers that cite. Page and box travel with every chunk into the answer.
- Built in. Document RAG runs on the same engine.
A folder of papers, LLM-ready.
The same parser runs inside .NET with LM-Kit.NET, or over REST, a command line and MCP with LM-Kit One.
LM-Kit One · terminal
# Every paper in ./papers as Markdown: LaTeX formulas, plots as tables, sections kept
lmkit run document-parsing --in ./papers --out ./markdown --effort High --output-format Markdown
LM-Kit.NET · PaperToMarkdown.cs
using LMKit.Document.Parsing; using var parser = new DocumentParser(ParsingEffort.High); ParsedDocument paper = parser.Parse("paper.pdf"); File.WriteAllText("paper.md", paper.ToMarkdown()); // for the LLM File.WriteAllText("paper.json", paper.ToJson()); // pages, boxes, confidence
Unpublished research stays in the lab.
Preprints, lab notebooks and patent drafts are intellectual property. The parser runs on your own machines, air-gapped if needed.
Edge
On a lab workstation
Embed the parser in a .NET tool on the machine where the papers already are.
Edge deploymentLocal
For the whole institute
Serve it from LM-Kit One to any language, notebook or agent on your network.
Deploy LM-Kit OneSovereign
Under your jurisdiction
No page leaves your infrastructure, and LM-Kit never receives a document.
SovereigntyFrequently asked questions.
How are equations returned?
Display formulas come back as LaTeX in the Markdown, JSON, HTML and DocLang output, so they can be rendered, searched or passed to an LLM without losing their structure.
Can it read the values in a plot?
Yes. Bar charts, line plots, pies and multi-panel figures are returned as the data table they plot, with legend and units, so a result shown only as a curve becomes values an index can find.
Does it handle two-column layouts?
Yes. Elements come back in the order a person reads them, with captions bound to their figures and footnotes and running heads kept apart from the body.
Can answers cite the paper precisely?
Every element carries its page, bounding box and confidence. Parsed papers feed Document RAG, which answers with the page and region behind each claim.
Scientific papers