A two-column research paper can look fine in a PDF viewer and turn into interleaved paragraphs when copied into a text file. A table may retain every number while losing the column headings that explain those numbers. Those errors carry through to a search index or an AI assistant's answer.
KolmoPDF processes technical PDFs into Markdown and translates PDFs while preserving their layout. It is intended for research papers, manuals and reports where equations, tables or page structure affect the meaning of the document.
Choose the output your task needs
For a knowledge base or a document-processing pipeline, use PDF to Markdown. The parser handles multi-column layouts, LaTeX formulas, tables and code blocks. Translation can run during parsing, which is useful when the downstream corpus needs a common language.
Markdown gives you a file that can be inspected, edited and split into retrieval chunks. A structured table can keep column headings alongside its values, and formulas can retain a LaTeX representation. Compare the output with the source on the pages that matter to your task.
For reading or sharing a translated document, use PDF translation. This produces a translated PDF with the original layout, with a bilingual output option for checking the translation against its source. Scanned pages are supported. This is a different output from translated Markdown: choose it when page composition matters to the reader.
Start with a document you can inspect
The homepage includes a public example using the 15-page Attention Is All You Need paper. You can view the source pages, rendered Markdown and translated outputs, or download the original PDF and parsed Markdown.
That example is a useful introduction to the output format. For your own project, select a representative document containing the structures you actually use. Compare a table with its source, read across a column transition and inspect an equation before processing a larger collection. A clean first page says little about a dense appendix.
Once the output is suitable, keep the original PDF alongside the Markdown. That gives a reviewer a direct way to check a retrieved passage or resolve an ambiguous symbol later.
Use the same service from an agent or application
The developer guide covers API access and agent setup. KolmoPDF also has an installable agent skill:
npx skills add komoai2026/kolmopdf-skill
Configure KOLMOPDF_API_KEY in the environment your agent uses. MCP is optional for that skill. The web interface, API and MCP use the same account credits, so a document tried in the browser can inform the processing workflow you later automate.
This is useful when PDF conversion is one step inside a research or ingestion task. An agent can request structured document text through the service; your application can then apply its own chunking, indexing or review process. Parsing quality and answer quality remain separate things to evaluate.
The web application also includes image-to-Markdown OCR, document-to-slides generation and format-conversion utilities. Start with the operation your workflow needs rather than putting every document through every available transformation.
Where KolmoPDF fits
The main advantage is support for technical document structure together with a choice of Markdown or a translated PDF. It is worth evaluating when plain extraction loses reading order, when equations need a usable representation, or when a translated report must remain recognizable page by page.
A short PDF with ordinary selectable text may already work with a simpler local extractor. Complex scans and unusual layouts need representative testing; no parser can guarantee perfect output for every PDF. KolmoPDF's standard service processes documents online, so use documents you are authorized to upload.
For background on the parsing problems, read the guides to multi-column reading order and table semantics in Markdown. To process a document, open PDF to Markdown; for application integration, use the API documentation.
Top comments (2)
Thanks for stopping by my post! I really liked your example of a table keeping all its numbers but losing the headers that give them meaning. It made me think about how a RAG system can retrieve the right numbers and still give the wrong answer if that context is lost during parsing. I'd be curious to see what the Markdown output looks like for a table that spans multiple pages.
When a table crosses a page break, standard PDF extractors usually emit two detached Markdown tables, and the second one starts immediately with data rows without column names.
In a layout-aware pipeline, the parser tracks column boundary coordinates across the boundary. If the column x-spans align and vertical text flow continues, it strips the repeated header row (and any running page footers like 'Page 14') from the second page and stitches the rows into a single continuous table:
For RAG chunking where the full table exceeds the token budget, the chunker splits by row groups and re-injects the original header block at the top of each fragment. That way, an embedding chunk containing only the second half still pairs every numeric cell with its column name.