Summary

Extraction that checks its own work.

The distinguishing move is the audit: most extractors hand back whatever they got and stay quiet about the pages they mangled, and that silent garbage is what poisons a RAG pipeline. pdfmux routes per page, re-extracts the failures with a stronger backend, and — crucially — flags what it still cannot read rather than dropping it. The `verify` path turns the same audit outward, certifying any other engine's output for silent drops. It runs fully locally by default; a Gemini key is only for a paid fallback on pages every local backend finds hard, and the README is candid that the certifier is lexical, catching missing and garbled content, not faithful paraphrase.

What it is

A PDF extraction tool with an MCP server, built so an AI agent gets clean text rather than silent gaps. It routes each page to the best-fitting extraction backend, then audits the result — catching blank pages, scrambled columns and broken tables — re-extracts the failures with a stronger backend, and marks what it still cannot read. A separate `verify` path audits any other extractor's output against the source PDF for pages that were silently dropped.

What you get

PDFs converted to Markdown, JSON or RAG-ready chunks — `convert_pdf`, `extract_streaming`, `batch_convert`
A PDF analysed and its metadata read before extraction — `analyze_pdf`, `get_pdf_metadata`
Structured extraction against a schema or preset (invoice, receipt, contract, resume, paper) — `extract_structured`
Per-page routing across built-in backends — PyMuPDF, OpenDataLoader, RapidOCR, Docling, Surya, Marker, Mistral OCR — with a BYOK LLM fallback
An audit that re-extracts pages the first backend got wrong, and flags what it still cannot read rather than dropping it
Certification of any extractor's output against the source PDF, returning a verdict, the silently dropped page numbers, and a coverage score
First-class Arabic, Persian, Urdu and Hebrew handling with automatic RTL reordering

Requirements

`pip install pdfmux` for digital PDFs; the README strongly recommends `pdfmux[ocr]` for any real batch, since almost every directory has at least one scan that returns empty text without OCR. Further extras add table extraction (`[tables]`), neural extraction (`[marker]`) and specific LLM providers. The MCP server runs with `pdfmux serve` (stdio or `--http`), or via the npm package `pdfmux-mcp` with `npx -y pdfmux-mcp`. Python 3.11+. The pyproject names the package `pdfmux` at 1.8.7. An optional `GEMINI_API_KEY` enables a paid Gemini fallback for low-confidence pages — pdfmux runs fully locally without it. MIT licensed; the certify path is patent-clean and reuses only the shipped audit layer.

Setup effort

One command — pip install pdfmux