Compare · Pandoc and MarkItDown
Webfold vs Pandoc and MarkItDown
Webfold completes 6.6 times the calls a second of Pandoc. It completes about 777 times those of MarkItDown. That is one process per document at 16 workers. Pandoc has no XLSX reader, so 80 of its 288 calls failed. This measures speed, not fidelity.
Pandoc is the standard document converter. MarkItDown is Microsoft's Python tool for models. We ran all three as command-line calls, one process per document. A second test scores fidelity against what Pandoc 3.1.3 reads from 352 files. Those scores are for superdocuments main, newer than the Webfold command line's pin.
Webfold on its own
How fast is Webfold alone?
Each job is measured on its own. The performance page states the figures once. We link them here.
- Convert Word, Excel, PowerPoint, OpenDocument and ebooks to Markdown: 27.7 µs a document, in process. 203,854 documents a second on 64 workers.
- Read PDFs, with OCR when the page is a picture: 5 to 12 ms a text file; 0.6 to 1.0 s a scanned page. Not measured at scale.
- Read LaTeX papers and BibTeX bibliographies: About 4 ms a LaTeX file. Not measured at scale.
Now the comparison
Webfold against Pandoc and MarkItDown
| Aspect | Webfold | Pandoc and MarkItDown |
|---|---|---|
| Calls per second, 16 workers | 2,173 | Pandoc 329; MarkItDown 2.8 |
| Seconds per document, MarkItDown | Under 0.01 s | About 2.3 s (starts a Python interpreter per file) |
| Excel (.xlsx) | Reads it. | Pandoc has no .xlsx input; those jobs errored. |
| Reads it, with an OCR fallback. | Pandoc has no PDF input. | |
| LaTeX, 16 fixtures | 2 disagreements, both one figure* environment. | Pandoc 3.1.3 is the reference. |
| Fidelity to Pandoc's reading, 352 documents in 7 formats | Mean 0.977 (0.764 before the Pandoc-compatibility work) | MarkItDown 0.658 on the same files and scorer; it does not read ODT, so its 34 ODT files count as failures |
| LaTeX, 118 documents | Mean fidelity 0.968 (0.857 before); 10 documents identical | MarkItDown 0.428; Pandoc is the reference |
| BibTeX to citation data, 179 entries | 179 of 179 identical to Pandoc's (0 of 179 before) | Pandoc is the reference |
| Formatted citations against pandoc --citeproc | 7,454 of 7,458 citations and 4,362 of 4,362 bibliography entries identical on a development set (125 documents, 6 styles); on 4 styles not used while developing, 5,079 of 5,080 and 2,214 of 2,952 after fixes (4,929 and 1,923 on the first run) | Pandoc is the reference; on the CSL project's 845 edge-case tests 529 match completely |
| BibTeX, 5 fixtures | Reads all five. | Pandoc converts two of the five. It stops on a parenthesis-delimited entry and on a missing comma, and runs out of memory on a cross-reference cycle. |
| Licence | Apache 2.0. | Pandoc GPL; MarkItDown MIT. |
Where is each tool ahead?
Where Pandoc and MarkItDown is ahead
- Pandoc reads dozens of formats. It writes them too. It has a mature citation processor. It has years of edge-case coverage.
- The fidelity scores take Pandoc's reading as the reference. Pandoc wrote the corpus's HTML files. It wrote its EPUB, RST or Markdown files too. That favours a reader that copies it. Deliberate differences are in the benchmark's disagreements file.
- Pandoc's citation processor has years of coverage. On the CSL project's edge-case tests, 316 of 845 differ from Pandoc. They are not explained one by one.
- MarkItDown has a Python-native API. It has a plugin ecosystem.
Where Webfold is ahead
- 6.6 times Pandoc's command-line throughput. About 777 times MarkItDown's. There is no interpreter or library start-up per file.
- One binary that also fetches. It also crawls. It also archives.
- PDF with OCR fallback. Excel. Pandoc does not read either.
How do I move from Pandoc and MarkItDown to Webfold?
| If you use | In Webfold |
|---|---|
| pandoc -t gfm file.docx | webfold convert file.docx |
| markitdown file.xlsx | webfold convert file.xlsx |
| pandoc -f latex | the LaTeX reader in superdocuments (library); the webfold command line does not route .tex files |
What is the evidence?
Each figure states its dataset. It states its limits.
FAQ
Frequently asked questions
Is Webfold's Markdown as good as Pandoc's?
Measured against Pandoc's own reading of the same files: mean fidelity 0.977 over 352 documents (DOCX 0.972, EPUB 0.985, HTML 0.990, LaTeX 0.968, ODT 0.948, reStructuredText 0.996, Markdown 1.000), up from 0.764 before the Pandoc-compatibility work. These scores are for superdocuments main, which is newer than the revision the Webfold command line pins. Fidelity is structure, features (headings, lists, tables, links, math, footnotes, citations) and words compared with Pandoc's reading; it says nothing about layout or about documents unlike the corpus. The throughput tracks measure speed and success, not fidelity.
Why is MarkItDown so far behind?
Most of the time is Python and library start-up for every file, which is exactly the cost a single binary avoids. A long-lived MarkItDown process would close much of the gap.
Try it on your own pages.
Run the command-line tool. Join the waitlist for the hosted API.