pdf-inspector: Half Your PDFs Never Needed OCR

Firecrawl open-sourced pdf-inspector, a Rust library that answers one question in 10 to 50 milliseconds: does this PDF actually need OCR? Their claim: about 54% of PDFs are text-based and don't. Classify locally, extract text with position and font info, convert to clean Markdown with headings, tables and code blocks intact, and only route the genuinely scanned stuff to your expensive OCR service. It jumped 1,700+ stars in a day to 8k total, MIT licensed, with bindings for Python, Node, browser WASM, and a CLI.

Why this is agent news: every document workflow, from RAG pipelines to agents reading contracts, invoices and papers, has been paying the OCR tax on every single PDF because nobody trusted a cheap way to tell the difference. Vision-model OCR costs real money and adds seconds per page. A sub-200ms local Rust pass that scores 0.875 on their extraction benchmark, top of the five competitors they tested, turns the default pipeline from "OCR everything" into "OCR the 46% that earns it". One dependency (lopdf), dual-mode table detection, CID fonts, RTL text, multi-column layouts.

Firecrawl keeps running the same play: find the boring chokepoint in the agent data-ingestion stack, write the fast version, open-source it. Boring chokepoints are where the compounding infrastructure value lives.

Repo: https://github.com/firecrawl/pdf-inspector