How fast, how complete, and how clean is anydoc, really? Rather than taking our word for it, here is the official head-to-head data published in the README.

Methodology

• Sample: 100 real documents covering all 14 formats (Word / PowerPoint / Excel / OpenDocument / RTF / EPUB / CSV / PDF)
• Contenders: anydoc, LibreOffice, unstructured, markitdown, pandoc, docling, and mammoth — seven in total
• Dimensions: completeness, structure, format fidelity, and cleanliness, each scored out of 100
• Speed: median conversion time per tool

Results

Tool	Formats	Median	Score	Completeness	Structure	Fidelity	Cleanliness
anydoc	14/14	4.4ms	81	87	79	78	81
libreoffice	12/14	1129.5ms	40	59	42	40	24
unstructured	8/14	572.9ms	58	63	76	59	51
markitdown	6/14	134.8ms	33	65	78	66	60
pandoc	5/14	102.1ms	34	56	74	57	56
docling	4/14	513.6ms	21	57	60	60	57
mammoth	1/14	52.5ms	8	70	84	71	75

How to read this table

Format coverage: 14/14 — the only full sweep. Every other tool has blind spots: mammoth only handles docx, pandoc and docling cover fewer than half. If your workflow sees any format at all, only anydoc gets you through on a single code path.

Speed: 4.4ms vs 134.8ms. The next-fastest converter, markitdown, is more than an order of magnitude slower; LibreOffice is 250× slower. In batch pipelines, real-time parsing, and agent calls, this gap is the difference between usable and painful.

Quality: 81 overall, leading every column. The 87 completeness score matters most — it means the least information loss, exactly what you care about when feeding Markdown to an LLM. Structure, fidelity, and cleanliness all beat the runner-up too.

Don't forget the operational cost. LibreOffice requires a system install and carries its own runtime; docling and unstructured drag in heavyweight models and dependency chains. anydoc is a small Rust library — a few MB inside your binary or WASM bundle.

Why anydoc is this fast

Speed is an architectural choice, not magic:

1. One model — every format parses into a shared Document model, so there is exactly one rendering path and no format-to-format re-conversion
2. Pure Rust — no interpreter overhead, no process boundaries, no model inference
3. Parses only what it needs — headers, footers, and page numbers are excluded by design, so no wasted work

Be clear about the limits

• Scanned PDFs are not supported: image-only PDFs need OCR and return Unsupported. Firecrawl Parse layers OCR onto the same conversion pipeline.
• The benchmark is one sample: 100 documents are representative but not the universe — extreme layouts should still be tested against your own files.
• Speed is a median: very large files (hundreds of MB of xlsx) take longer, though the relative advantage over comparable tools holds.

Want to run it yourself?

The official repo ships a full benchmark harness in bench/ — clone it and run it against your own document set for the most honest numbers.

Want the head-to-head with markitdown?

No comparison of the two biggest open-source converters is complete without it — markitdown vs anydoc: which one converts documents better?

Contributors: souvc