papero-extract

papero-extract --versionpapero-extract extract file [-o path] [-f format] [-p pages] [--fast] [options]papero-extract batch input -o directory [options]papero-extract serve [--host address] [--port port]

papero-extract rebuilds the structure of a PDF: reading order, tables, formulas, figures, and a bounding box for every block. It is the command-line entry point of papero (the papero-extract Python package). Two engines run on the same file. A layout engine on PDFium reads glyphs, rules, and images and rebuilds columns and tables from geometry. Apache Tika adds metadata, tagged headings, OCR, and non-PDF formats such as DOCX, PPTX, XLSX, EPUB, and HTML.extract handles one file. With no -f, the format follows the -o extension (.md, .txt, .json, .html, .csv) and otherwise is Markdown. csv writes tables only. --images crops figures, tables, and formulas to PNG; when -o is set, those files land in an images/ directory next to the output. --fast skips the layout engine and returns cleaned text from Tika.batch walks a directory (default pattern **\*.pdf) and writes one Markdown and one JSON file per document, a chunks.jsonl file, a manifest, and a fidelity report under fidelity/**. The report compares the extraction to the words PDFium reads on each page. It is a guide to documents worth reviewing, not a score against a ground-truth corpus.serve starts the REST API and the bundled browser app. The API extra must be installed (pip install 'papero-extract[api]'). POST /v1/extract accepts an uploaded file. Interactive docs are at /docs.

extract file

Extract one document. Writes to stdout unless -o is set.

Output file. CSV is written as UTF-8 with a BOM. Other formats are UTF-8.

One of markdown, text, json, html, csv. Default: inferred from -o, otherwise markdown.

Page selection such as 1-3,5,10-.

Password for an encrypted PDF.

Crop figures, tables, and formulas to PNG.

Crop resolution. The default 2 is 144 dpi.

Skip table detection.

Skip formula detection.

When to OCR. The default auto OCRs pages that look scanned. force OCRs every page. off never does.

Tesseract language string. The default is por+eng.

Run the PDFium layout engine only, with no Java Tika server.

Mark the start of each page in Markdown.

Text only, via Tika. Fastest path. Layout, tables, and formulas are not rebuilt.

With --fast, skip all text cleanup.

With --fast, keep headers and footers.

With --fast, join words split by a hyphen at a line break.

Worker processes. For extract, the default is the CPU count. For serve, workers are off unless this is set (exported as PTE_WORKERS).

Suppress the stderr summary. Warnings are still printed.

input is a directory (walked recursively) or a single file. -o is required and is the dataset directory.

Files to include. Default **\*.pdf**.

Per-document outputs, comma-separated. Default markdown,json.

Maximum chunk size in characters. Default 1500. Chunks break on headings and keep tables whole.

HTTP API and browser app. Defaults to host 0.0.0.0 and port 8000, or the PORT environment variable when it is set.

Bind address.

Listen port.

Print the package version and exit.

serve fails until the API extra is installed: pip install 'papero-extract[api]'. OCR and non-PDF formats need a working Tika and Tesseract setup. --no-tika drops both and keeps only the geometry engine.Formula reconstruction is glyph-based. Fractions, roots, exponents, and indices become LaTeX. Matrices and aligned systems come out linear. A cropped image of the formula is the reliable copy when --images is on. Borderless tables with very narrow column gaps can be read as paragraphs.Word (.docx) and Excel export exist in the browser app. The CLI formats are Markdown, text, JSON, HTML, and CSV. Scanned pages are not readable with --fast. That path warns when it finds little text.

pdftotext(1), pdftohtml(1), mutool(1), tesseract(1), pandoc(1)

Source code · Homepage