What it does Zerox converts PDFs, DOCX files, and images into a sequence of page images, then asks a vision model (GPT-4o, Claude 3, Gemini, etc.) to transcribe each one into Markdown. It aggregates the results and hands back structured text with optional per-page metadata like token counts and completion time. There’s also a JSON Schema extraction mode if you need structured data instead of raw Markdown.

The interesting bit The maintainFormat option chains pages sequentially: page 1’s Markdown gets fed into the prompt for page 2, which helps preserve tables that span page breaks. The trade-off is speed — concurrency drops to 1. It’s a blunt but effective hack for a genuinely hard layout problem.

Key highlights

• Supports OpenAI, Azure, AWS Bedrock, and Google Gemini with provider-specific credential passing
• Node and Python SDKs, though feature parity is uneven (Node has schema extraction and orientation correction; Python has custom system prompts and Vertex AI)
• Optional per-page structured extraction with a separate model/provider from the OCR step
• Configurable concurrency, DPI, image compression, and temp directory management
• Requires system dependencies: graphicsmagick/ghostscript for Node, poppler for Python

Caveats

• The Node and Python versions diverge significantly; check the feature matrix before picking one
• No pricing or latency benchmarks are provided, and token counts from the example suggest multi-page documents could get expensive quickly
• README mentions Tesseract workers in the API but doesn’t explain when they activate vs. vision-model fallback

Verdict Worth a look if you’re already paying for vision-model API access and your documents have complex layouts that defeat conventional OCR. Skip it if you need deterministic, offline, or low-cost extraction — this is fundamentally a cloud-API wrapper with image preprocessing glue.