Most PDF parsers treat a page as a flat stream of text, which is why tables collapse and reading order scrambles the moment a layout goes two-column. This service starts from the opposite premise: work out the geometry first — what's a title, a footnote, a table, a formula — and only then extract text in the right order.
What Sets It Apart
- Two interchangeable engines for the same job: a Vision Grid Transformer (originally from Alibaba's research group) that reads the whole page image for ~0.96 F1 on PubLayNet, and a Poppler + LightGBM ensemble that runs ~0.42s/page on CPU. So you trade accuracy against hardware instead of being locked to a GPU.
- Eleven distinct segment labels — caption, footnote, formula, list item, page header/footer, picture, section header, table, text, title — not just "text vs. image". Downstream RAG or extraction can filter or weight regions by type.
- It goes past detection into reconstruction: reading-order sorting, Tesseract OCR in 150+ languages, tables to HTML, formulas to LaTeX, plus Markdown/HTML export. One container turns a scanned PDF into structured, machine-usable output.
Who It's For + Tradeoffs
Great fit if you're building a document-ingestion or RAG pipeline and need clean, typed, correctly ordered segments from messy real-world PDFs — especially scanned or multi-column ones — and want to run it as a self-hosted Docker microservice, picking the engine per workload. Look elsewhere if you only need plain text from clean digital PDFs (a lightweight extractor is enough), or if you can't host a GPU and the CPU LightGBM accuracy falls short for your domain.
