Markitdown vs Tika: What Each Text Extractor Misses, Measured

A text extractor can hand back clean text and still drop the comment in the margin, the chart's title or the page header, or let a deleted paragraph back in. We hid a marker in every place 17 file formats can hold text, ran the same files through FileConcat, Apache Tika and Microsoft's markitdown, and counted what came back.

FileConcat 2.6, in every format but ebooks. Of the other two, Tika reads more, and markitdown reads no OpenDocument file at all.

* markitdown printed these files as their own source, MIME parts and base64 for email, control words for RTF. That is not text a model can use, so it scores zero. How it scores counted as text is under "How we ran it". The ranking does not change.

Each score is the share of what a reader of the file sees that came back. A construct found in most real files counts more than one found in few, so a spreadsheet's page header outweighs a chart's axis title.

How it was scored: a marker in each construct. A visible construct counts when its marker comes back, a hidden or deleted one when it stays out.

markitdown leaves out comments, page headers and link addresses. Tika reads more of them, but also prints what the author deleted or hid.

Tika's miss is the riskier kind. A deleted sentence in a Word file comes back run into the sentence it was cut from, with nothing to mark it, and a model reads it as the document's text. Across every hidden or deleted piece in the set, FileConcat keeps nearly nine in ten out and Tika fewer than half.

Unweighted, every construct counts the same however rare it is, and Tika moves just ahead of FileConcat 2.3.

What changed from 2.3: comments with their authors, charts and SmartArt, page headers and footers, form field values, link addresses, calendar invites and bounce reasons, and deleted or hidden text kept out.

Text in an old encoding, logs padded with zero bytes, and compiled files posing as text. FileConcat reads or names all of them, and every other tool misses a few.

The misses take three shapes. Tika returns nothing for a binary lock file, prints compiled Lua and a DRM-locked ebook chapter as garbage, and garbles Turkish saved in an old Windows encoding. markitdown stops with an error on two build logs, and for a text file with binary data after it returns the word None. FileConcat 2.3 misread most of the legacy-encoded files.

How it reads: the first bytes decide whether a file is a document, an archive, text or binary. Text in an unknown encoding goes to chardetng, the detector Firefox uses.

On three constructs. One is a choice, two are open.

1. Ebook authors. markitdown names the book's author. FileConcat leaves document authors out on purpose, the same way in every format. Three in four real ebooks carry one, which costs about a tenth of the ebook score and is why markitdown leads that row.
2. A PDF line drawn invisible. Tika and markitdown print a line set to invisible on an ordinary page. FileConcat drops it. It is rare in real PDFs.
3. A PDF comment flagged hidden. It still comes back. None of the real PDFs we sampled had one.

Every format FileConcat reads, and how, is on the formats page. Your files are not uploaded to us, the whole thing runs in your browser.

Each probe file went through each tool once, on one machine, one process per file with a 300 second limit. FileConcat ran its web reading chain in Node, with HTML through the web app's own reader. Tika ran as java -jar tika-app-3.3.2.jar --text. markitdown ran as its command line, installed with uvx --from 'markitdown[all]==0.1.5'.

The rule was written before any total was read. A visible construct counts when every one of its markers is in the output. A hidden or deleted one counts when its marker is absent and the same file gave output with at least one other marker, so a tool that reads nothing earns nothing for keeping things out. A crash, an empty output or an unsupported format scores zero. Where the map gives a construct two parts with different answers, a name to hide and a text to show, each part is held to its own.

Raw source counts as not read. markitdown prints email and RTF as the file's own text, and Tika prints a notebook as its JSON. Counted as text instead, they score:

Counted as text, markitdown's RTF scores 89.5 weighted, above FileConcat's 85.6, because control words leave the marker strings intact. That is the case the rule exists for.

We also judged FileConcat 2.6's output by hand, construct by construct. The automatic rule disagrees on 45 of 648. It credits 6 that the hand judgment does not, on layout and reading order, which marker presence cannot see. It misses 39 that the hand judgment counts as handled, most of them document properties FileConcat leaves out by design. Every column in this post uses the automatic rule, so the comparison is like for like. One automatic miss is the rule's own: a formula with a TeX source is printed as its TeX, and the probe's TeX held only a marker.

Hidden and deleted text: 130 constructs carry a "keep it out" answer. FileConcat 2.6 keeps 115 out, FileConcat 2.3 97, Tika 62 and markitdown 45. markitdown's figure is low partly because a format it does not read earns nothing for keeping things out.

FileConcat's losses, sized on the real files: the ebook author is in 76.0% of 200 ebooks and worth 11.6 points of the ebook score. An invisible line of PDF text is on 0.5% of 195 PDFs, worth 0.1 points. No sampled PDF had a hidden comment.

Percent of the format's marked constructs. * raw source, counted as not read.

The weighted table covers the 11 formats we have real files for. A construct no real file of the format carried has no weight and leaves that format's weighted score.

The 45 cases are FileConcat's own test set for files that are hard to tell or hard to decode, built in memory from a seeded generator so no binary is committed. Extractors pass a text case when 90% of its words come back with tags stripped, and a binary case only when they say they cannot read it. Empty output or garbage fails. The empty file passes on empty output. file, git and ripgrep pass when their text-or-binary call is right.

FileConcat ran each tree's own byte router and text classifier on every file, with the encoding detector where the tree has one (2.6 yes, 2.3 no). Without it, 2.6 passes 38: the seven legacy-encoding cases need the detector. ripgrep searches a file named on its command line even when it is binary, so its binary misses follow its design for explicit paths.

Published comparisons of these tools measure speed, or score whole-document quality against a reference. We found none that counts which constructs inside a file come back, which deleted or hidden text leaks, or which hard files a tool names instead of garbling.

The text-quality side of FileConcat on public benchmarks, including olmOCR-bench for PDFs, is in the browser extraction benchmark.

• Home field. The probes were written for FileConcat's own coverage map, and FileConcat was fixed against that map between 2.3 and 2.6. The hard files are its own test set. The one counterweight is small: the xls and ppt probes were written with Apache POI, the library Tika reads those formats with, and Tika still scores 70 and 88 on them against FileConcat's 100.
• Presence, not order. A construct counts when its marker comes back. Reading order, wrapping and whether a comment is labelled as one are not scored, which is why the hand judgment disagrees on 45 constructs.
• Weights for 11 formats. The weighted table rests on 65 to 200 real files per format, from one Common Crawl sample. Outlook, the 97-2003 formats, notebooks and subtitles have no weights.
• Word 97 .doc is out. Its constructs carry no markers, so it is in neither table.
• One run. One machine, one pass, no timing.
• Repositories not counted. A fix in this version for files whose binary data starts late changed no verdict across fresh clones of 60 public repositories, 39,229 files, since fresh clones hold almost no compiled files. It is not in these numbers.

Everything but the Tika jar is in the repository: the probes, the scorers, the parsed constructs and the real-file counts behind the weights.

D=packages/cli/scripts/coverage/compare
bash packages/cli/scripts/coverage/run.sh . out/fc-HEAD
bash packages/cli/scripts/coverage/run.sh ../fc-v2.3.0 out/fc-v2.3.0
TIKA_JAR=tika-app-3.3.2.jar bash $D/run-tools.sh tika packages/cli/scripts/coverage/probes out/tika
bash $D/run-tools.sh markitdown packages/cli/scripts/coverage/probes out/markitdown
RULE=2 python3 -I $D/score.py $D/rows.json packages/cli/scripts/coverage/probes out $D/prevalence-2026-10-16 report.json
python3 -I $D/tables.py report.json

The hard files, the comparison figures and the version worktree are in packages/cli/scripts/coverage/README.md.

Is markitdown or Tika better for text extraction? Tika read more on our probes: 67.7% of marked constructs against markitdown's 35.0%, and 72.9 against 41.5 weighted by real files. markitdown led on ebooks. Tika also printed deleted and hidden text that markitdown kept out.

Does markitdown read OpenDocument files? No. markitdown 0.1.5 stops on odt, ods and odp with "No converter attempted a conversion". It also does not read PowerPoint 97 files.

Does Apache Tika include tracked changes in the text? Yes. Deleted text in Word and RTF files comes back in the body, run into the surrounding sentence, with nothing to mark it as deleted. FileConcat keeps it out.

Can I get Word comments and PowerPoint notes into the text? FileConcat 2.6 reads comments with their authors in Word, PowerPoint, Excel, OpenDocument and RTF, and the comments on a PDF. In our probes Tika returned a Word comment without its author and missed comments in PowerPoint's current format. markitdown returned neither.

Do these tools upload my files? Tika and markitdown run on your machine. FileConcat runs in your browser: your files are not uploaded to us. What you then paste into an assistant is your own choice, but preparing the file uploads nothing.

• Apache Tika. tika.apache.org
• Microsoft. markitdown. github.com/microsoft/markitdown
• Apache POI test data. github.com/apache/poi
• Li, Z. et al. "READoc: A Unified Benchmark for Realistic Document Structured Extraction." ACL 2025 Findings. arXiv:2409.05137
• Kreuzberg. "Kreuzberg v4.3.0 and comparative benchmarks." 2026. dev.to
• Firecrawl. anydoc. github.com/firecrawl/anydoc
• Mozilla. chardetng. github.com/hsivonen/chardetng
• Common Crawl. commoncrawl.org