kreuzberg-dev/kreuzberg v4.3.3

on GitHub

What's New in v4.3.3

PaddleOCR Multi-Language Support (#388)

• 106+ language support via 12 script families: PaddleOCR recognition models now cover english, chinese (simplified+traditional+japanese), latin, korean, east slavic (cyrillic), thai, greek, arabic, devanagari, tamil, telugu, and kannada script families.
• Per-family recognition model architecture: Shared detection/classification models with per-family recognition models and dictionaries, downloaded on demand from HuggingFace.
• Engine pool for concurrent multi-language OCR: Replaced single-engine architecture with a per-family engine pool, enabling concurrent OCR across different languages.
• Backend-agnostic --ocr-language CLI flag: Works with all OCR backends (tesseract, paddle-ocr, easyocr).
• SHA256 checksum verification: All model downloads verified against embedded checksums.

Centralized Image OCR Processing

• Shared process_images_with_ocr function for all document extractors (DOCX, PPTX, Jupyter, Markdown).

Jupyter Notebook Image Extraction

• Base64 image decoding from notebook cell outputs with OCR support.

Markdown Data URI Image Extraction

• Data URI image decoding with OCR support for embedded images.

DOCX Full Extraction Pipeline (#387)

• DocumentStructure generation, pages field population, OCR on embedded images.
• Typed metadata fields, style-based heading detection, markdown formatting.
• Performance optimizations: eliminated 3x code duplication, removed unnecessary clones.

Fixed

• LaTeX zero-arg command handling preventing silent text loss.
• Structured data is_text_field false positives from substring matching.
• PaddleOCR CrnnNet recognition height fixed from 48 to 32 pixels.

See CHANGELOG.md for full details.