AXA’s Document Parser Rebuilds PDFs, but It’s Abandoned

What it does

Parsr ingests PDFs, Word documents, emails, and images, then emits structured data in JSON, Markdown, CSV, or plain text. It attempts to reconstruct logical document hierarchy—headings, paragraphs, tables, lists, headers, footers, and tables of contents—rather than dumping a flat stream of words. The system is packaged as a Dockerized API with an optional web GUI and a Python client.

The interesting bit

Instead of treating documents as bags of words, Parsr orchestrates a small army of specialized engines—Tesseract for OCR, Camelot for tables, pdfminer.six and PDF.js for text extraction—to regenerate reading order and structure. That makes it more of a document-understanding glue layer than a simple format converter, which is both its charm and its limitation.

Key highlights

• Exports to JSON, Markdown, CSV/Pandas DataFrames, or TXT
• Detects headings, tables, lists, page numbers, headers/footers, and links
• Ships as a Dockerized API with an optional browser-based GUI
• Python client (parsr-client) exists, though the README offers no API surface detail
• Core is Apache 2.0, but optional dependencies include AGPL (MuPDF) and GPL (Pandoc) components

Caveats

• Abandoned. The README posts a clear warning: the project is unmaintained, security patches are not applied, and the team suggests LiteParse as a replacement.
• License friction. Optional dependencies pull in AGPL and GPL code, which can complicate redistribution if you enable them.
• Accuracy is outsourced. Because Parsr delegates to external tools for OCR, table detection, and text extraction, its output quality is bounded by those upstream projects.

Verdict

A useful reference architecture for anyone building a multi-format document pipeline, but choose a maintained alternative for production workloads. Avoid if you need vendor support or guaranteed security updates.