Stop OCRing Every PDF: Route It First with pdf-inspector
OCR is often the most expensive and slowest step in a document-ingestion pipeline. The frustrating part is that many PDFs already contain usable text, yet a naive pipeline sends every document through OCR anyway. pdf-inspector takes a better approach: classify first, extract native text when possible, and route only the pages that actually need OCR. The routing pattern The core decision is simple: PDF arrives ↓ Classify the document and its pages ├─ native text available → extract locally → Markdown └─ text missing/broken → route those pages to OCR That small decision can remove a large amount of unnecessary OCR work from RAG ingestion, invoice processing, research-paper parsing, and document search. The library classifies PDFs as: TextBased Scanned ImageBased Mixed It also returns a confidence score and the specific pages that need OCR. A 40-page report with one scanned appendix does not have to become a 40-page OCR job. Quick start in Python Install the package: pip install pdf-inspector Then process a PDF: import pdf_inspector result = pdf_inspector.process_pdf(“document.pdf”) print(result.pdf_type) print(result.pages_needing_ocr) print(result.markdown) For selective OCR, the native package also exposes an OCR-aware pipeline: ocr_result = pdf_inspector.process_pdf_with_ocr(“document.pdf”) print(ocr_result.pages_routed_to_ocr) The OCR runtime remains separate and is only touched when a page is routed to it. That keeps the default extraction path lightweight. Node.js and browser support The same idea is available for Node.js: npm install @firecrawl/pdf-inspector import { readFileSync } from “fs”; import { processPdf } from “@firecrawl/pdf-inspector”; const pdf = readFileSync(“document.pdf”); const result = processPdf(pdf); console.log(result.pdfType); console.log(result.markdown); There is also a WebAssembly package for running the Rust parser locally in a browser or Web Worker: npm install @firecrawl/pdf-inspector-wasm This is useful when documents should not be uploaded to a parsing service just to determine whether they contain native text. What the extractor preserves Classification is only half the project. For text-based PDFs, the extractor attempts to preserve structure such as: headings derived from font-size tiers bold and italic text numbered and bulleted lists code blocks detected from monospace fonts tables detected from drawing rectangles and text alignment multi-column reading order links, page breaks, captions, and common font encodings The output is Markdown, which makes the library convenient for search indexing and LLM/RAG pipelines. How classification works At a high level, the detector inspects PDF content streams for text operators such as Tj and TJ, and image operators such as Do. It can scan all pages, stop early, sample a large document, or inspect a caller-provided page set. This is a routing signal, not a promise that every PDF will be perfectly parsed. PDFs with broken encodings, text converted to vector paths, or extremely complex layouts may still need OCR or a specialized parser. The library explicitly reports encoding problems so callers can fall back instead of silently accepting bad text. About the benchmark numbers The project publishes a reproducible benchmark against a 200-document corpus. Its July 2026 results report strong reading-order and table scores as well as fast local processing. Those are project-published measurements on specified hardware—not a universal latency guarantee—so benchmark your own document mix before committing to production thresholds. The more durable takeaway is architectural: OCR should be a fallback chosen per page, not the default chosen per file. A practical production rule A conservative router might look like this: result = pdf_inspector.process_pdf(“document.pdf”) if result.pdf_type == “text_based” and result.confidence >= 0.95: store_markdown(result.markdown) else: send_pages_to_ocr(result.pages_needing_ocr) Your threshold should depend on the cost of a false positive. A casual knowledge base can tolerate more extraction noise than a legal or financial workflow. If your pipeline currently OCRs every incoming PDF, classification-first routing is a small change with a clear operational payoff. The longer version and implementation notes are available on ToolGenix.