kreuzberg-dev/kreuzberg v4.0.0-rc.6

on GitHub

[4.0.0-rc.6] - 2025-12-07

Release Candidate 6 - FFI Core Feature & CI/Build Improvements

New Features

FFI Bindings:

• Added core feature for kreuzberg-ffi without embeddings support
  • Provides lightweight FFI build option excluding ONNX Runtime dependency
  • Enables Windows MinGW compatibility for Go bindings
  • Includes HTML processing and all document extraction features
  • Use --no-default-features --features core for MinGW builds

Bug Fixes

ODT Extraction:

• Fixed ODT table extraction producing duplicate content
  • Table cells were being extracted twice: once as markdown tables (correct) and again as raw paragraphs (incorrect)
  • Root cause: XML traversal using .descendants() included nested table cell content as document-level text
  • Solution: Changed to only process direct children of <office:text> element, isolating table content
  • Impact: ODT extraction now produces clean output without cell duplication
• Enhanced ODT metadata extraction to match Office Open XML capabilities
  • Added comprehensive metadata extraction from meta.xml (OpenDocument standard)
  • New OdtProperties struct supports all OpenDocument metadata fields
  • Extracts: title, subject, creator, initial-creator, keywords, description, dates, language
  • Document statistics: page count, word count, character count, paragraph count, table count, image count
  • Metadata extraction now consistent between ODT, DOCX, XLSX, and PPTX formats
  • Impact: ODT files now provide rich metadata comparable to other Office formats

Go Bindings:

• Fixed Windows MinGW builds by disabling embeddings feature
  • Windows ONNX Runtime only provides MSVC .lib files incompatible with MinGW
  • Go bindings on Windows now use core feature (no embeddings)
  • Full features (including embeddings) remain available on Linux, macOS, and Windows MSVC
• Fixed test execution to use test_documents instead of .kreuzberg cache
  • Ensures reproducible test runs without relying on user cache directory
  • Improves CI/CD reliability and test isolation

CI/CD Infrastructure:

• Upgraded upload-artifact from v4 to v5 for compatibility with download-artifact@v6
  • Fixes artifact version mismatch causing benchmark and CI failures
  • Affects 10 workflow files with 42 total changes
  • Resolves "artifact not found" errors in multi-job workflows
• Fixed RUSTFLAGS handling in setup-onnx-runtime action
  • Now appends to existing RUSTFLAGS instead of overwriting
  • Preserves -C target-feature=+crt-static for Windows GNU builds
• Fixed Go Windows CI artifact download path causing linker failures
  • Changed download path from target to . to prevent double-nesting (target/target/...)
  • Linker can now find libkreuzberg_ffi.dll at correct location
  • Added debug logging to show directory structure after artifact download
• Aligned all workflows to Java 24
  • Updated from Java 23 to 24 across all CI and publish workflows
  • Resolves "release version 25 not supported" compilation errors
  • Affects ci-validate, ci-java, publish, and benchmarks workflows

Ruby Bindings:

• Fixed rb-sys links conflict in gem build
  • Removed rb-sys vendoring, now uses version 0.9.119 from crates.io
  • Resolves Cargo error: "package rb-sys links to native library rb, but it conflicts with previous package"
  • Allows Cargo to unify rb-sys dependency across magnus and kreuzberg-rb

C# E2E Tests:

• Fixed OCR tests failing with empty content
  • Added render_config_expression function to C# E2E generator
  • Tests now pass proper OCR config JSON instead of null
  • Regenerated all C# tests with tesseract backend configuration
• Fixed metadata array contains assertion for single value in array
  • Extended ValueContains method to handle value-in-array case
  • Fixes sheet_names metadata assertions in Excel tests

Python Bindings:

• Fixed missing format_type in text extraction metadata
  • TypstExtractor and LatexExtractor incorrectly claimed text/plain MIME type
  • Removed text/plain from both extractors' supported types
  • PlainTextExtractor now correctly handles text/plain with proper TextMetadata
  • Metadata now includes format_type, line_count, word_count, character_count
  • Added unit test for Metadata serialization to verify format field flattening