markitdown vs pdf-inspector - PDF to Markdown Compared

markitdown and pdf-inspector are two of the most frequently compared open-source "PDF to Markdown" tools — one from Microsoft, one from Firecrawl. Both free, both promising clean text for LLMs. But run them on a batch of real PDFs and the gap is an order of magnitude.

This article is a factual breakdown based on the official opendataloader-bench (200 real PDFs, five engines, OCR off everywhere). Data first, conclusions second.

The key metrics at a glance

Metric	pdf-inspector	markitdown	Notes
Maintainer	Firecrawl (open-source Rust)	Microsoft (open-source Python)	Different ecosystems
Core positioning	Classify-first parsing engine	Universal format converter	One specialized, one broad
Type detection	10–50ms verdict + confidence + per-page signals	None	The key lever on OCR costs
Overall score (0–1)	0.875	0.589	#1 of five engines
Reading order	0.915	0.844	Multi-column quality
Tables (TEDS)	0.814	0.273	The largest single gap
Headings	0.788	0.000	markitdown barely recovers headings
Corpus time (200 docs)	0.470s	16.165s	34×+ faster
Bindings	Rust / Node / Python / WASM / CLI	Python (CLI & .NET ports exist)	pdf-inspector ships five ways
Runs in the browser	WASM, no cross-origin isolation needed	Requires Python runtime	pdf-inspector works pure-frontend
OCR strategy	Selective — only pages that need it	Via plugins	Different philosophies

Source: the pdf-inspector repo benchmark (opendataloader-bench corpus, 200 real PDFs, refreshed July 2026).

The short version

PDF is your main input, output feeds LLMs, tables and multi-column matter → pick pdf-inspector. Most accurate overall, tables by a landslide, 30× faster — plus classify-first routing so you only pay OCR when a document actually needs it.
You're in a Python project with wildly mixed inputs (URLs, PPTX, XLSX, audio) and want the plugin ecosystem → markitdown still works, but its PDF numbers (tables 0.273, headings 0.000) mean you'll probably need a dedicated PDF parser as backup anyway.

Two different tools

What markitdown is

Microsoft's open-source converter (github.com/microsoft/markitdown) — a Python library + CLI that turns many formats into Markdown for LLM consumption. Strengths:

Mature ecosystem: high community energy since its 2024 launch, familiar to Python developers, rich third-party integrations
Broad format reach: beyond Office files it covers images, audio transcription, EPUB, ZIP, and more
Extensible: plugin system supports custom behaviors like LLM image descriptions

Its weaknesses on PDFs (per the benchmark):

Overall score 0.589 — last among the five engines
Headings 0.000: essentially no heading-level recovery, so the Markdown comes out structurally flat
Tables TEDS 0.273: table structure is mostly guesswork
16.165s for the corpus — over 34× slower than pdf-inspector
Hard dependency on Python; can't run in the browser

What pdf-inspector is

Firecrawl's engine (github.com/firecrawl/pdf-inspector) — a Rust parser with five bindings, and the underlying engine of Firecrawl's hybrid OCR pipeline. Strengths:

Classify first, extract second: TextBased / Scanned / ImageBased / Mixed in 10–50ms with confidence and per-page OCR needs, keeping costs predictable
Quality first: overall 0.875, reading order 0.915, table TEDS 0.814 — all #1
Millisecond-fast: native-text PDFs finish end-to-end locally within 200ms; the whole corpus runs in 0.470s
Layout-aware: multi-column reconstruction, CID/Type0 fonts decoded via ToUnicode CMaps (CJK-friendly)
Flexible deployment: npm / PyPI / crates.io / browser WASM; the WASM build never uploads a byte

Its boundaries:

PDF only — Word, Excel, and friends are out of scope
No OCR in the default build; scanned documents need the native selective-OCR mode or an external service

Scenario-based recommendation

Your scenario	Pick	Why
Bulk PDF ingestion for RAG / fine-tuning	pdf-inspector	Speed + tables + structure win directly improves corpus quality
Mixed inputs (Office/web/audio all present)	markitdown + fallback	Use markitdown as glue; route heavy PDF work to a dedicated engine
High-concurrency server-side parsing	pdf-inspector	Async APIs run on the libuv thread pool without blocking the loop
Pure frontend, files must stay local	pdf-inspector	WASM build needs no COOP/COEP headers
Controlling OCR spend	pdf-inspector	Per-page routing signals + selective OCR; ~54% of PDFs skip OCR entirely

Transparent data

Every number above comes from the benchmark published in the pdf-inspector repo (OCR disabled, local engines compared), with public methodology and artifacts. A representative benchmark still isn't the whole universe — test on your own documents before committing.