firecrawl/pdf-inspector

pdf‑inspector – Fast, smart PDF classification & extraction

What it does

• Quickly decides whether a PDF is text‑based, scanned, image‑based or a mix (10‑50 ms) and gives a confidence score.
• For text‑based PDFs it extracts the text with layout information (font, X/Y coordinates, rotation) and turns the document into clean Markdown.
• Detects headings, lists, code blocks, tables (both rectangle‑based and heuristic), bold/italic, URLs, page breaks, etc.
• Optional selective OCR: only pages that lack extractable text are sent to a local PP‑OCRv6 model, avoiding full‑document OCR latency.
• Available as a native Rust crate, a CLI (pdf2md / detect-pdf), Python and Node.js packages, and a WebAssembly build for browsers.

Why it matters

• About half of PDFs on the web are already text‑based; running OCR on all of them wastes CPU and money. pdf‑inspector lets you route only the truly scanned pages to OCR, cutting cost and latency.
• The library keeps the PDF parsed once, sharing that representation between detection and extraction, which speeds up the whole pipeline.
• Benchmarks on a 200‑PDF corpus show it out‑performs comparable tools in overall quality, reading order, table handling, and speed (0.47 s for the whole set vs. several seconds for alternatives).

How to use it

• Rust: cargo add pdf-inspector; call process_pdf("file.pdf") to get pdf_type and optional Markdown.
• Python: pip install pdf-inspector; pdf_inspector.process_pdf("file.pdf").
• Node: npm install @firecrawl/pdf-inspector; processPdf(buffer).
• Browser: import the WASM package and call processPdf(uint8Array).
• CLI: pdf2md file.pdf for Markdown, detect-pdf file.pdf for just the classification.

Key technical bits

• Classification samples a few pages, looks for text (Tj/TJ) vs. image (Do) operators.
• Extraction walks the PDF content stream, builds TextItems with font metrics and positions, then groups them into lines, columns, and tables.
• Table detection combines rectangle detection from drawing commands with heuristic alignment of text.
• Markdown conversion uses font‑size ratios for heading levels and font‑name patterns for bold/italic.

When to pick it

• You need fast, local processing of large batches of PDFs (reports, research papers, invoices, legal docs).
• You want structured Markdown with tables and proper reading order without paying for a cloud OCR service.
• You run in environments where Rust, Python, Node, or the browser are available, and you can install the optional PDFium/ONNX Runtime libraries for OCR when needed.

License: MIT.