// the take
A Rust PDF toolkit from Firecrawl that classifies a PDF as text-based, scanned, image-based, or mixed and converts the text-based case straight to structured Markdown without OCR, with Python, Node, and WASM bindings plus a CLI. It's aimed at pipelines that process PDFs at volume and need to decide cheaply which documents actually need OCR before paying for it.
Parses the document once and shares it between detection and extraction instead of loading twice. Table detection is dual-mode — rectangle-based from the PDF's own drawing operators plus a heuristic alignment-based fallback — which covers the much harder case of tables with no ruled lines, not just the easy ones. Benchmarks are run against real competitors (liteparse, opendataloader, pymupdf4llm, markitdown) on a public corpus with a reproducible harness, not just self-reported numbers. Per-page OCR routing with a versioned JSON envelope (confidence, provenance, hosted-fallback hints) means callers can audit why a page got routed instead of trusting a black box, and it correctly handles CID/Type0 fonts and RTL text, which is where most naive PDF extractors fall apart.
OCR support requires separately installing and version-pinning PDFium and ONNX Runtime native libraries plus model files outside the Python/Node packages — that's real operational overhead for anyone building from Rust or the CLI with the OCR feature. Classification defaults to sampling 8 pages (Sample(8)), so a large mixed document with scanned sections buried in the middle can be misclassified as text-based and silently skip OCR for those pages — it's a documented speed/accuracy trade-off, but it's the kind of thing that bites people in production. Heading and structure detection is font-size-ratio heuristics, which is inherently fragile against PDFs with unconventional or heavily-styled typography. No mention of how it handles maliciously crafted or deeply corrupted PDFs beyond a SECURITY.md pointer, which matters if this sits in front of user-uploaded files.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →