Text Extraction in the Browser: 9 File Types Benchmarked
Measured on October 7, 2026 with FileConcat 2.1 and kept as measured. FileConcat has changed since. What it reads today is listed in the formats list. The latest run of these benchmarks is on the benchmarks page.
FileConcat reads every file in the browser it was dropped into, with no upload and no GPU. Version 2.1 replaced most of its readers. We ran public benchmarks on both versions.
01 of 12
How good is FileConcat 2.1 at each file type?
Within a few points of the best server tools where a file stores its own text, behind the GPU readers where text has to be recognised from a scan or a voice.
| File type | 2.1 | 2.0 | Best other reader | Measure |
|---|---|---|---|---|
| 40.0 | 25.7 | Chandra, GPU: 83.1 | olmOCR-bench | |
| Saved web page | 93.7 | 8.8 | rs-trafilatura: 97.0 | article body F1 |
| Archive | 25 of 26 | 13 of 26 | 7-Zip wasm: 25 of 26 | text files exact |
| PowerPoint (ppt) | 63.4 | not read | Tika ceiling: 73.1 | F1 vs Tika |
| Excel (xlsx) | 74.1 | 55.1 | F1 vs Tika | |
| RTF | 71.9 | 62.5 | Tika ceiling: 76.0 | F1 vs Tika |
| Kindle books | 99.6 | not read | F1 vs source | |
| Video subtitles | 0% | not read | word error | |
| Video speech | 12.7% / 32.4% | not read | word error |
Higher is better, except word error. Speech is English / Turkish. The Tika ceiling is how far Tika agrees with itself across a LibreOffice conversion of the same files. Word and Excel 97-2003 files scored the same in both versions.
02 of 12
PDF: how close does a browser get to a GPU?
Half way. 2.1 passes about as many of olmOCR-bench's tests as LiteParse run alone, the best other reader without a GPU that we ran, and none of them writes math.
Share of olmOCR-bench's 8,413 tests passed, mean of its 8 parts. GPU and API rows: allenai's published table (Chandra self-reported). Marker, CPU: its own README, self-reported. LiteParse, anydoc and FileConcat 2.0: run by us in Node under scorer f7cfe4c. FileConcat 2.1: in the browser, same scorer, 2026-10-07
| Part | Tests | FileConcat 2.1 | LiteParse | FileConcat 2.0 | olmOCR 2 |
|---|---|---|---|---|---|
| arXiv math | 2,927 | 0.0 | 0.0 | 0.0 | 83.0 |
| old math scans | 458 | 0.0 | 0.0 | 0.0 | 82.3 |
| tables | 1,022 | 53.0 | 53.1 | 31.5 | 84.9 |
| old scans | 526 | 17.3 | 13.3 | 13.3 | 47.7 |
| header/footer | 760 | 55.0 | 55.5 | 39.2 | 96.1 |
| multi-column | 884 | 64.0 | 65.5 | 20.5 | 83.7 |
| tiny text | 442 | 30.3 | 23.8 | 14.0 | 81.9 |
| baseline | 1,394 | 100.0 | 99.9 | 87.2 | 99.7 |
| overall | 8,413 | 40.0 | 38.9 | 25.7 | 82.4 |
Most of the jump from 2.0 is columns and tables, the rest headers, tiny text and the baseline: 2.1 keeps a two-column page in reading order and a table's cells under their headings. A quarter of the score is equations written as LaTeX, which a text layer cannot give, so every reader without a vision model scores zero there.
How it reads: LiteParse on pages with a text layer, OCR on pages without one.
03 of 12
Saved web pages: is the article all that comes back?
Yes, for a page saved from Chrome. 2.0 bundled the whole HTML source, markup included.
Word 4-gram F1 x 100 against the benchmark's article bodies, its own evaluate.py at commit 4a3bc97. Other readers: the published outputs in the benchmark repository. FileConcat 2.1: article body, links as their text. FileConcat 2.0: the same pages bundled as HTML source, which is what 2.0 did with every saved page
2.1 lands one point under Mozilla Readability, the engine it runs, and three under the best server extractor. A page without a saved-page mark FileConcat recognizes, from Chrome or SingleFile, still goes the 2.0 way.
How it reads: Chrome's "saved from" mark, then Mozilla Readability, then Markdown.
04 of 12
Office files: does an old format still read?
Every Office format we tested reads in 2.1 except Word 95 and older and dotx templates, close to how far Tika agrees with itself across a conversion.
| Format | Files | 2.0 | 2.1 | Tika ceiling |
|---|---|---|---|---|
| ppt | 52 | not read | 63.4 | 73.1 |
| rtf | 63 | 62.5 | 71.9 | 76.0 |
| doc | 63 | 75.0 | 75.0 | 80.4 |
| xls | 164 | 66.8 | 66.8 | 71.9 |
| xlsx | 163 | 55.1 | 74.1 | |
| xlsb | 6 | 0 | 66.1 |
Word 4-gram F1 x 100 against Apache Tika 3.3.2. The ceiling is Tika scored against itself on a LibreOffice-converted copy. Office files have no public answer key, so agreement with a mature reader stands in for one.
How it reads: SheetJS for workbooks, a reader of its own for Word 97-2003, anydoc for ppt and rtf, officeparser for the rest. Each has a second reader for when the first returns nothing.
05 of 12
Archives: does every file inside come back?
All but one, level with 7-Zip run on its own and ahead of libarchive.
26 archives built from known files, one per case readers get wrong: zip methods past deflate, zip64, non-ASCII names, four tar dialects, gz, bz2, xz, 7z. 7-Zip wasm, libarchive and FileConcat 2.0: run by us in Node. FileConcat 2.1: in the browser, 2026-10-07
The one miss is a zip inside a zip, which is not opened. On libarchive's own rar and 7z test files, damaged and encrypted ones included, every file finished without a crash or a hang.
How it reads: zip, tar and gz in JavaScript, bz2, xz, rar and 7z through 7-Zip compiled to WebAssembly.
06 of 12
Ebooks and video: what did 2.0 not read at all?
Kindle books and video. Both now read, a subtitle track word for word.
| What was read | Score |
|---|---|
| epub, mobi, azw3 against Project Gutenberg's text | F1 98.8, 99.6, 99.6 |
| video subtitle track, mp4, mkv, webm | 0% word error |
| video speech, English / Turkish | 12.7% / 32.4% word error |
| on-screen text in video without burned-in subtitles, English / Turkish | F1 82.7 / 81.8 |
Speech is the weakest row. English runs on Moonshine base, a 66 MB model. Every other language runs on Whisper small, 244 MB, and the Turkish sample is two videos.
How it reads: the subtitle track on drop, speech and on-screen text only when you ask.
07 of 12
Where does FileConcat still lose?
On three kinds of file, and each has a way around it.
- Scans and math. The GPU readers at the top of the PDF chart are vision models with billions of parameters. The English OCR model FileConcat downloads is 3 MB. For a scan you need exact, run it through one and drop the text it returns.
- Web pages saved by Firefox. Save the page again from Chrome, or paste the article into a text file.
- Speech that is not English. Keep a subtitle track in the file when there is one. It is read exactly and costs nothing.
Every format FileConcat reads, and how, is on the formats page.
Try it on your own files
Drop a scan, an Office file or a video and read the first lines of the result.
08 of 12
How we ran it
Every file was dropped into FileConcat 2.1's production build in headless Chromium, the way a visitor drops it, and the text was taken from the downloaded Plain bundle. Each benchmark's own scorer did the scoring.
Chromium ran with an English or Turkish language setting, matching each video's speech, since that setting picks the speech and OCR model. Documents went in batches of up to 20 from one folder. A PDF that came back without text was dropped again alone, because FileConcat 2.1 read at most three scans per drop by itself. Archives and videos went one per drop.
The first Turkish run came back empty: the first transcription in a session ignored the browser's language and used the English model. That was fixed before the Turkish rerun. The English rows were not rerun.
Web pages are scored three ways, because the bundle adds lines the benchmark counts as noise.
| FileConcat 2.1 output | F1 | Precision | Recall |
|---|---|---|---|
| article body, links as their text | 0.937 | 0.901 | 0.977 |
| title and source lines kept, links as their text | 0.908 | 0.848 | 0.977 |
| as written, title and source lines kept, links as Markdown | 0.851 | 0.762 | 0.963 |
| the same pages without Chrome's "saved from" mark (the 2.0 path) | 0.088 | 0.047 | 0.873 |
Office files were measured in two sets: 587 modern files converted to the old formats by LibreOffice, scored against Tika on the original, and Apache POI's own test files. The table above uses the converted set for doc, xls, ppt and rtf and POI's files for xlsx and xlsb. On POI's files, 2.1 scored doc 74.5 (99 files), xls 74.8 (241), ppt 78.6 (71), docx 66.5 (64), pptx 55.9 (55) and xlsm 52.5 (7). Files with fewer than 20 reference words were not scored.
None of the published extraction benchmarks we list here runs its readers inside a browser, and most are written by the vendor of the tool that wins them. We found no other measurement of browser extraction either, for any format.
GPU vision models, OCR APIs and pipeline tools
Chandra 0.1.0 83.1, olmOCR 2 82.4, PaddleOCR-VL 80.0, Marker 1.10.1 76.1, Mistral OCR API 72.0
CPU text-layer readers
LiteParse no OCR 39.6, pdf-inspector 33.7, opendataloader 32.5, markitdown 28.7
Marker against other pipelines
Marker fast no-OCR (CPU) 43.6, LiteParse no OCR 20.4
34 open-source and commercial article extractors
rs_trafilatura and AutoExtract 0.970, trafilatura 2.0.0 0.958, readability.js 0.947
anydoc, markitdown, docling, unstructured, pandoc
DOCX anydoc 88, markitdown 71, docling 71. Run natively, not in a browser
Moonshine against Whisper of similar size
moonshine-base 3.23 / 8.18 WER, whisper-base.en 4.25 / 10.35
The sources contradict each other where it matters here. LiteParse scores 39.6 on olmOCR-bench in its own README and 20.4 in Marker's, with OCR off in both. We ran every reader on our side under the same scorer, so a gap between two of our rows is not a gap between two harnesses.
09 of 12
Limitations
- One browser. Chromium only. Firefox and Safari were not measured.
- No timing. Runs were on a desktop, so no speed figure is published.
- 2.0 ran in Node. Its rows are its extraction code run outside the browser. For 2.1, seven of the nine Office figures with a Node run matched the browser to the third decimal, and the other two were within 0.012. The two settings read the same.
- Small speech samples. Turkish is two videos. Its 32.4% is a direction, not a leaderboard figure.
- Agreement is not truth. Where a reader keeps or drops words Tika does not, such as a header, footer or comment, those words lower its score, and LibreOffice wrote every converted file.
- Archives favour 7-Zip. 7-Zip wrote the 7z variants and reads them in FileConcat. rar has no keyed set, since only RARLAB's own tools write it.
- Weak and small: a Word 6 file scored 0, since the Word reader declines Word 6 and Word 95, and two dotx templates scored 0.9.
- Not measured: layout beyond the benchmarks' own tests, email, encrypted files.
10 of 12
Reproduce it
Build the web app and serve it, then drop a corpus through it.
pnpm build
cd apps/web && pnpm vite preview --port 4719 --strictPort
node packages/cli/scripts/measure-live-build.mjs http://localhost:4719/ <corpus dir> out.json
node packages/cli/scripts/measure-live-build.mjs http://localhost:4719/ <pdfs dir> out.json --olmocr <candidate dir>
node packages/cli/scripts/measure-live-build.mjs http://localhost:4719/ <video dir> out.json --transcribe
Score with packages/cli/scripts/score-against-reference.py (Office, ebooks),
score-video.py (video), olmOCR-bench's python -m olmocr.bench.benchmark at commit f7cfe4c
(PDF) and the article benchmark's evaluate.py (web pages). The archive and video test sets are
written by build-archive-corpus.py and build-video-corpus.py in the same folder, with a fixed
seed. packages/cli/scripts/bench/rerun-browser.sh reruns FileConcat's PDF, web-page and archive rows: it adds
Chrome's "saved from" mark to the benchmark pages, reshapes the bundle into the reader-only view
scored for the web-page rows, and scores archive output against the known files.
11 of 12
Frequently asked questions
Can a browser extract text from a PDF without uploading it? Yes. FileConcat 2.1 reads the text layer in the browser and runs OCR on pages that have none. On 1,403 benchmark pages it scored 40.0, against 38.9 for LiteParse run alone and 82.4 for olmOCR 2 on a GPU.
How accurate is in-browser OCR? Weak. On olmOCR-bench's old scans 2.1 passed 17.3% of tests, against 13.3% for LiteParse and FileConcat 2.0, which run no OCR, and it scored zero on math. The GPU readers that lead the benchmark are far larger than anything FileConcat downloads.
Can I get text out of a saved web page? Yes, if it was saved from Chrome. On 181 benchmark pages the article body matched at 0.937 F1, one point under Mozilla Readability. The bundle as copied, with its title line and Markdown links, scores 0.851. A page saved by Firefox is bundled as HTML source.
Does it read old Word, Excel and PowerPoint files? doc, xls and ppt from Word 97 onward, plus rtf. Agreement with Apache Tika was 0.63 to 0.79, close to how far Tika agrees with itself across a conversion. The Word reader declines Word 6 and Word 95 files.
Can it transcribe a video? On request. A subtitle track is read exactly and at once. Without one, speech is transcribed in the browser after a one-time model download: 66 MB for English, 244 MB for any other language.
12 of 12
References
- Poznanski, J. et al. "olmOCR 2: Unit Test Rewards for Document OCR." Allen Institute for AI, 2025. arXiv:2510.19817
- Zyte. Article extraction benchmark. github.com/scrapinghub/article-extraction-benchmark
- Jeffries, N. et al. "Moonshine: Speech Recognition for Live Transcription and Voice Commands." 2024. arXiv:2410.15608
- Hugging Face. Open ASR Leaderboard. huggingface.co
- LlamaIndex. LiteParse. github.com/run-llama/liteparse
- Datalab. Marker. github.com/datalab-to/marker
- Apache Tika. tika.apache.org
Tip
What gets lost inside a document when it becomes text, table by table and footnote by footnote, is its own study: what gets lost converting documents to text.