Kreuzberg Development Roadmap

[Goldziher]

The following is my current thinking about the roadmap. It shouldn't be seen as set in stone, and the time boxing is also tentative and might be a bit over ambitious.

Please give your feedback!

Note About Versioning

Kreuzberg follows a rapid development cycle with regular major version releases. Why? Because major versions (1.x, 2.x etc.) rather than pre-1.0 versioning (0.1.11 etc.) allows us to evolve the library's interfaces confidently while being explicit about breaking changes. And we do and will always follow SemVer.

Current: Version 2.x

Core Functionality

• Unified async/sync API for document text extraction
• Support for PDF, images, Office documents, and markup formats
• OCR capabilities via Tesseract integration
• Text extraction and metadata extraction via Pandoc
• Efficient batch processing

Version 3.x (Q2 2025)

Extensibility

Architecture Update:

• Support for creating and using custom extractors for any file format
• Capability to override existing extractors
• Pre-processing, validation, and post-processing hooks
• Extended metadata extraction

Enhanced Document Structure

Optional Features (available via extra install groups):

• Multiple OCR backends (Paddle OCR, EasyOCR, etc.) with Tesseract becoming optional
• Table extraction and representation (awaiting release as V3.1)

Possible path forward (TBD):

Version 4.x (Q3 2025)

• Automatic language detection
• Entity/keyword extraction

Model-Based Processing

Optional Vision Model Integration:

• Structured text extraction using open source vision models (QWEN 2.5, Phi 3 Vision, etc.)
• Plug-and-play support for both CPU and GPU (via HF transformers or ONNX)
• Custom prompting with structured output generation (similar to Pydantic for document extraction)

Optional Specialized OCR:

• Support for advanced OCR models (TrOCR, Donut, etc.)
• Auto-finetuning capabilities for improved accuracy with user data
• Lightweight deployment options for serverless environments

Optional Heuristics:

• Model-based heuristics for automatic pipeline optimization
• Automatic document type detection and processing selection
• Result validation and quality assessment
• Parameter optimization through automated feedback

Version 5.x (Q4 2025)

Integration & Ecosystem

Optional Enterprise Integrations:

• Connectors for major cloud document platforms:
  • Azure Document Intelligence
  • AWS Textract
  • Google Cloud Document AI
  • NVIDIA Document Understanding
• User-provided credential management
• Standardized response format using Kreuzberg's data types
• Integration with Kreuzberg's intelligent processing heuristics

TBD...