v4.0.0-rc.5

Release Candidate 5 - macOS Binary Fix

Bug Fixes

Build System:

• Fixed macOS CLI binary missing libpdfium.dylib (dyld error at runtime)
  • Build script now correctly copies libpdfium.dylib to target-specific directory when using --target flag
  • Resolves: dyld: Library not loaded: @rpath/libpdfium.dylib
  • Impact: macOS CLI binary now functional in releases

Note: rc.4 workflow fixes for Python, Node, Ruby, and Maven were committed after rc.4 tag, causing those packages not to publish. All fixes are now present for rc.5.

v4.0.0-rc.3

Release Candidate 3 - ONNX Runtime Bundling & Test Fixes

Breaking Changes

None - this is a patch release over rc.2

Bug Fixes

ONNX Runtime Distribution

• Switched from ort-load-dynamic to ort-download-binaries feature
• ONNX Runtime binaries are now bundled at build time
• Fixes embeddings feature across ALL release variants (npm, PyPI, RubyGems, Maven, crates.io)
• Users no longer need to install onnxruntime as a system dependency

Node.js Bindings

• Fixed TypeScript build configuration (tsconfig.json rootDir)
• Fixed all test imports from src/ to dist/ (461 tests now passing)
• Fixed flaky forceOcr: true test (OCR backend isolation issue)
• Fixed test cleanup to prevent mutex errors

Documentation

• Added CONTRIBUTING.md symlink to fix broken GitHub link

Known Issues

• A cosmetic mutex error may appear at the very end of Node.js test runs during C++ static destruction
• This occurs after all Rust code completes and can be safely ignored
• Will be resolved when fastembed upgrades to ort >= 2.0.0

Test Results

• Rust: All tests passing
• Node.js: 461/461 tests passing (100%)
• Python, Ruby, Java, Go: TBD via CI

What Changed Since rc.2

• ONNX Runtime is now properly bundled
• Node.js package fully functional and tested
• All test isolation issues resolved

v4.0.0-rc.4

Release Candidate 4 - Critical CI/CD and Build Fixes

Bug Fixes

CI/CD Workflow Fixes:

• Fixed RubyGems action version (v2 doesn't exist, now using v1.0.0)
• Fixed pnpm workspace configuration (replaced invalid --cwd flag with -C)
• Fixed Docker environment variables (undefined $LD_LIBRARY_PATH in Dockerfiles)
• Fixed Maven credentials timing (env vars now available when setup-java generates settings.xml)
• Fixed Maven GPG configuration (modernized arguments to --pinentry-mode=loopback format)
• Removed release notes update job from publish workflow (not needed)

Core Library Fixes:

• Fixed Tesseract OCR test failure (corrected API call ordering: set_image before set_source_resolution)
• Fixed Go Windows CGO linking (build Rust FFI with x86_64-pc-windows-gnu target for MinGW compatibility)

Testing:

• All 24 Tesseract tests now pass (was 23/24 in rc.3)
• Go bindings now build successfully on Windows

v4.0.0-rc.2

Release Candidate 2 - C# Support & Infrastructure Improvements

Breaking Changes

TypeScript/Node.js Package Restructuring:

• NPM package renamed from kreuzberg to @kreuzberg/node (scoped package)
• Platform-specific packages now use @kreuzberg/{platform} naming scheme
• TypeScript source consolidated into crates/kreuzberg-node (merged with native bindings)
• Migration: Replace import { ... } from 'kreuzberg' with import { ... } from '@kreuzberg/node'
• See Migration Guide for details

New Features

C#/.NET Bindings:

• Complete C# bindings using .NET 9+ Foreign Function & Memory API
• Native FFI bridge via kreuzberg-ffi C library
• Supports .NET 9+ on Linux, macOS, and Windows
• Package: Kreuzberg on NuGet
• Full feature parity with other language bindings

Documentation Improvements:

• New v3 documentation site with MkDocs
• Comprehensive multi-language code examples for all 7 supported languages
• API reference documentation for all bindings
• Migration guides and tutorials

Bug Fixes

CI/CD Fixes:

• Fixed all CI workflow failures (ONNX Runtime, Maven dependencies, path triggers)
• Fixed benchmark harness configuration validation (single-file mode)
• Added ONNX Runtime installation to all relevant CI workflows
• Updated Maven plugins to latest compatible versions with enforcer plugin

Core Library Fixes:

• Added lock poisoning recovery to embeddings model cache
• Improved Tesseract tessdata path detection for Linux systems
• Fixed Python async test configuration (pytest-asyncio)
• Resolved Rust formatting issues with edition 2024 let-chain syntax

Benchmark Harness Improvements:

• Added comprehensive adapter registration diagnostics
• Fixed Tesseract benchmark path resolution
• Improved Python async benchmark output with performance metrics
• Added missing `--max-concurrent` parameter validation

Code Quality:

• Fixed Python linting issues (ruff complexity, mypy type parameters)
• Resolved all clippy warnings
• Fixed C# generated file formatting
• Improved error handling across all bindings

Developer Experience

Build System:

• Go FFI library and Ruby native extension build fixes
• Improved build reproducibility
• Better error messages during compilation

Testing:

• 268+ Rust tests passing
• 13/13 Java tests passing
• 32/32 benchmark harness tests passing
• Fixed 27 Python async test failures
• Unblocked 250+ tests across all language bindings