Biblicus

Make your documents usable by your assistant, then decide later how you will search and retrieve them.

Optical Character Recognition (OCR)

Traditional OCR extractors for text recognition from images and scanned documents.

OCR Extractors

Benchmarking

Overview

OCR extractors use computer vision to recognize text in images, scanned documents, and visual content. They are ideal for:

Available Extractors

ocr-rapidocr

RapidOCR provides fast ONNX-based text recognition with:

Installation: pip install biblicus[ocr]

Best for: General-purpose OCR, scanned documents, mixed text/image content

ocr-paddleocr-vl

PaddleOCR vision-language model provides:

Installation: pip install biblicus[paddleocr]

Best for: Complex documents, tables, multi-column layouts, CJK text

OCR vs VLM Document Understanding

When to Use OCR

When to Use VLM

For advanced document understanding with layout preservation, use VLM extractors:

VLM extractors provide:

Choosing an Extractor

Use Case Recommended Extractor Notes
English scanned docs ocr-rapidocr Fast, lightweight
Chinese/CJK documents ocr-paddleocr-vl Excellent CJK support
Tables and complex layouts docling-granite VLM approach
Simple screenshots ocr-rapidocr Quick results
Academic papers with equations docling-granite Equation recognition

Common Patterns

Fallback Chain

Try VLM first, fall back to OCR:

extractor_id: select-text
config:
  extractors:
    - docling-smol
    - ocr-rapidocr

Multi-Strategy Selection

Use longest output from multiple OCR approaches:

extractor_id: select-longest-text
config:
  extractors:
    - ocr-rapidocr
    - ocr-paddleocr-vl

Document Type Routing

Use smart overrides for different document types:

extractor_id: select-smart-override
config:
  default_extractor: ocr-rapidocr
  overrides:
    - media_type_pattern: "image/.*"
      extractor: ocr-rapidocr
    - media_type_pattern: "application/pdf"
      extractor: docling-smol

Performance Considerations

RapidOCR

PaddleOCR VL

VLM Alternatives

For best accuracy with complex documents, consider VLM extractors which offer:

See Also