kreuzberg-dev/kreuzberg v4.0.0-rc.14

on GitHub

Added

• Comprehensive test suites for all language bindings:
  • Python: 34 tests covering type verification, batch APIs, byte extraction, MIME detection, OCR, all file types
  • Node.js: 79 tests for type verification, batch APIs, MIME detection, configuration, error handling
  • Ruby: 55 RSpec tests for batch APIs, byte extraction, type verification, configuration handling
  • Java: 85 JUnit 5 tests with FFM API memory management, concurrency, all file format support
  • WASM: 79 tests with performance validation, large document handling, concurrent operations
  • Go: 63 table-driven tests with context support, error wrapping with errors.Is(), all file types
  • All test suites verify batch extraction APIs (sync/async), type safety, result structure validation

Fixed

• NuGet publish workflow reliability: Replaced NuGet/login OIDC-based authentication with direct API key approach
  • Issue: NuGet/login action could fail with 401 errors due to OIDC token context limitations (see NuGet/login#6)
  • Solution: Removed NuGet/login step and pass API key directly via NUGET_AUTH_TOKEN environment variable and --api-key parameter
  • Impact: More reliable C# package publishing without dependency on OIDC token exchange
  • Requires: NUGET_API_KEY secret to be configured in GitHub repository settings
• LibreOffice installation in Docker full image: Updated LibreOffice from 25.8.2 to 25.8.4
  • Version 25.8.2 download URLs were no longer available on DocumentFoundation servers
  • Updated to latest stable release 25.8.4 (released Dec 18, 2025)
  • Verified working for Office document extraction (DOCX, XLSX, ODT)
  • Tested on both x86_64 and aarch64 architectures
• Python IDE type completions: Fixed missing type hints in IDE autocomplete
  • Root cause: _internal_bindings.pyi stub file was not being included in wheel distribution
  • Solution: Added ensure_stub_file() function in packages/python/build.py to verify and include stub file in all build outputs
  • Impact: Full autocomplete now works for all 67 public APIs, type checkers can find definitions, mypy strict mode compatible
• Ruby gem native extension compilation: Fixed vendoring of Rust crates during build
  • Added automatic vendoring task to packages/ruby/Rakefile that runs before compilation
  • Ensures vendor/kreuzberg, vendor/kreuzberg-ffi, and vendor/kreuzberg-tesseract are properly copied and version-updated before building native extension
• Python ExtractionResult.pages type hints: Fixed missing type definition in PyO3 stub file
  • Root cause: _internal_bindings.pyi was missing pages field declaration in ExtractionResult class
  • Added pages: list[PageContent] | None attribute and PageContent TypedDict definition
  • Impact: IDEs now properly show autocomplete for result.pages, type checkers recognize the attribute
  • Fixes TypeError: 'NoneType' object is not iterable confusion when users iterate without checking for None