Zerox — Zero-Shot PDF OCR for AI Pipelines

What it is

Zerox is a Python library that extracts text from PDFs by converting each page to an image and then using vision-capable LLMs (GPT-4o, Claude, etc.) as the OCR engine. Unlike traditional OCR tools that require trained models for specific fonts and layouts, Zerox leverages the visual understanding of large language models to read any document format without training.

Data engineers processing scanned documents, researchers extracting text from academic papers, and developers building document processing pipelines use Zerox when traditional OCR produces poor results on complex layouts, tables, or handwritten content.

How it saves time or tokens

Traditional OCR pipelines require installing Tesseract, training custom models for specific document types, and writing post-processing logic to clean up OCR errors. Zerox replaces the entire pipeline with a single function call. Vision models handle complex layouts, tables, and multi-column documents that trip up conventional OCR. The output is clean markdown rather than raw text, reducing downstream parsing work.

How to use

1. Install Zerox:

pip install py-zerox

1. Extract text from a PDF:

from pyzerox import zerox
import asyncio

async def main():
    result = await zerox(
        file_path='report.pdf',
        model='gpt-4o-mini',
    )
    for page in result.pages:
        print(page.content)

asyncio.run(main())

1. The output is clean markdown for each page, ready for further processing or LLM consumption.

Example

from pyzerox import zerox
import asyncio

async def extract_with_claude():
    result = await zerox(
        file_path='financial_report.pdf',
        model='claude-3-5-sonnet-20241022',
        custom_system_prompt='Extract all text preserving table structure as markdown tables.',
    )
    # Each page returns clean markdown
    for i, page in enumerate(result.pages):
        print(f'--- Page {i+1} ---')
        print(page.content)
    
    # Save all pages to a single file
    with open('extracted.md', 'w') as f:
        for page in result.pages:
            f.write(page.content + '\n\n')

asyncio.run(extract_with_claude())

Related on TokRepo

• Document Processing Tools -- explore tools for PDF and document handling
• AI Tools for Research -- discover tools for academic and data research workflows

Common pitfalls

• Vision model API calls cost more than traditional OCR. For large documents (100+ pages), estimate API costs before processing. GPT-4o-mini is cheaper but less accurate than GPT-4o on complex layouts.
• Zerox converts each page to an image before sending to the model. High-resolution settings produce better results but increase API costs and processing time.
• The library is async by default. Wrap calls in asyncio.run() for synchronous usage, or integrate into an existing async application.