• Added output_format, result_format, elements, and djot_content fields to ExtractionResult
• Created proper PyChunk pyclass with attribute access (chunk.content) instead of raw dicts

• Unified output format: merged consolidated.json + aggregated.json into single results.json (schema v2.0.0)
• Added F1 token-based quality scoring with ground truth support
• Added OCR coverage for docling, unstructured, tika, mineru
• Naming normalization: strip -sync/-async suffixes
• Safety: eliminated unsafe set_var, added NaN sanitization, bounds checks, input validation
• Fixed zero-duration throughput inflation in batch results

See CHANGELOG.md for full details.

• Missing OutputFormat/ResultFormat exports with Python 3.10 compatibility
• .pyi stub alignment with Rust core (missing elements, Element, BoundingBox, PageHierarchy types)

• Config alignment with Rust core (field names, defaults, removed phantom parameters)
• Serialization test fixes and PHPStan compliance

• Missing elements field in NAPI-RS bindings
• outputFormat and resultFormat now correctly passed through config normalizer
• Serialization test import path fix

• asMap() null handling fix: absent config sections no longer incorrectly deserialized as default objects

• Enum serialization and test exception alignment

• Windows CI fix and e2e-generator support

• Bun runtime support

• Benchmark CI artifact size reduced from ~1.5GB to essential files only
• PageContent field parity across all language bindings

• Added Element, ElementType, BoundingBox, and ElementMetadata types to the TypeScript API surface
• Fixed batchExtractFile → batchExtractFiles export name

• Added #[serde(default)] to KeywordConfig fields for partial config deserialization

• Added JsonSerializable attributes for Element types (element_based result format)

• Removed deprecated Success field references

• Derived Jason.Encoder for ExtractionConfig struct

See CHANGELOG.md for full details.

See CHANGELOG.md for details.

See CHANGELOG.md for details.

• API parity: Added ExtractionConfig.new/0 and new/1 constructors for consistent struct creation
• Chunk field alignment: Changed text field to content for API parity with Rust core

• Error type fix: File-not-found errors now throw KreuzbergIOException instead of KreuzbergValidationException
  • Aligns with Rust error handling where file access issues are I/O errors

• Test alignment: Removed references to deprecated WithEmbedding() API and Chunking.Embedding field
• Test fixes: Updated config_comprehensive_test, config_result_test, embeddings_test, memory_safety_test

• ExtractionConfig expansion: Added embedding() and imagePreprocessing() builder methods
• Default value alignment: Fixed test assertions to expect enableQualityProcessing=true (matches Rust default)

• Rubocop compliance: Fixed Style/EmptyClassDefinition offenses in api_proxy.rb, cli_proxy.rb, mcp_proxy.rb

• Strict API parity enforcement: All 9 language bindings now have exact 1:1 field parity with Rust core
  • Verification script (scripts/verify_api_parity.py) now runs in STRICT mode, failing on ANY field differences
  • Added to ci-validate.yaml workflow to prevent future API drift

• ExtractionConfig alignment: Removed 5 non-canonical fields and fixed defaults
  • Removed: embedding, extractImages, extractTables, preserveFormatting, outputEncoding
  • Fixed defaults: useCache → true, enableQualityProcessing → true, maxConcurrentExtractions → null
  • Updated ExtractionConfigBuilder to match canonical API
  • All 16 fields now match Rust canonical source exactly

• ExtractionResult alignment: Removed Success field (not in Rust canonical)
• PageInfo alignment: Removed Visible and ContentType fields (not in Rust canonical)
• Updated 14 test files to remove references to removed fields

• Default value fix: Changed enable_quality_processing default from false to true to match Rust

• Default value fix: Changed enableQualityProcessing default from false to true to match Rust

• Type exports cleanup: Removed non-existent type exports (EmbeddingConfig, EmbeddingModelType, HierarchyConfig, ImagePreprocessingConfig) from index.ts

• Hex package compilation: Fixed force_build: true causing production installs to fail (#333)
  • Changed to force_build: Mix.env() in [:test, :dev] to only build from source in development
  • Production installs now correctly use precompiled NIF binaries from GitHub releases

• Tesseract OCR plugin initialization: Fixed "OCR backend 'tesseract' not registered" error in published Docker images
• Embeddings plugin initialization: Fixed "Failed to initialize embedding model" error

• JSON error responses: Added custom JsonApi extractor for consistent JSON error responses
• OpenAPI schema improvements: Enhanced schema validation constraints
• Chunking validation: Added validation that overlap must be less than max_characters
• Embed validation: Added validation that all text entries must be non-empty strings
• Default embedding model: EmbeddingConfig.model now defaults to "balanced" preset

• Schemathesis API contract testing: Added schemathesis to Docker CI workflow

• XLSX OOM with Excel Solver files: Fixed out-of-memory issue when processing XLSX files with sparse data at extreme cell positions (#331)

Patch Release: API Parity Fixes and CI Reliability Improvements

This patch release fixes API validation issues, adds missing format aliases, and improves backward compatibility across all language bindings.

• PPTX image page numbers: Fixed reversed page numbers when extracting images from PPTX files (#329)
  • Images on slide 1 were incorrectly reported with page_number=2 due to unsorted slide paths from presentation.xml.rels
  • Now sorts slide paths after parsing to ensure correct ordering regardless of XML element order
• Plugin registry error logging: Added comprehensive error logging for silent plugin failures (#328)
  • OCR registry now logs errors and warnings when plugins fail to initialize
  • Extractor registry logs plugin load failures for troubleshooting
  • PostProcessor registry tracks plugin status changes
  • Validator registry records plugin validation errors
  • New `startup_validation.rs` module provides plugin status verification
  • Server startup logs all active plugins and their initialization status (fixes Kubernetes deployment visibility)
• Output format validation: Extended `VALID_OUTPUT_FORMATS` to include all valid aliases (`plain`, `text`, `markdown`, `md`, `djot`, `html`)
• Error type consistency: `validate_file_exists()` now returns `Io` error instead of `Validation` error for file-not-found cases
• C# pre-commit hooks: Added dotnet restore to format/lint check tasks to fix failures in clean environments

• Format constants: Added `OutputFormatText` and `OutputFormatMd` as aliases for `plain` and `markdown`
• Documentation: Fixed default format comment (default is `plain`, not `markdown`)

• Format validation: Added `text` and `md` aliases to `validate_output_format` function
• Config validation: Updated error messages to list all valid format options

• CLI backward compatibility: `extract` and `detect` methods now accept both positional and keyword arguments
• Config field naming: Renamed `image_extraction` to `images` (canonical name) with backward-compatible alias
• Spec fixes: Updated test expectations to match actual implementation behavior

• Config field naming: Renamed fields to canonical names (`images`, `pages`, `pdfOptions`, `postprocessor`, `tokenReduction`)
• API parity: Added missing `postprocessor` and `tokenReduction` fields

• API parity: Added `getImages()` and `images()` builder methods as aliases for `getImageExtraction()`

• TypeScript types: Added `outputFormat`, `resultFormat`, and `htmlOptions` to `ExtractionConfig` interface

• Case sensitivity: Fixed tests to use lowercase format strings (`plain`, `unified`, `element_based`)
• API usage: Updated to use module-level functions (`config_to_json`, `config_merge`) instead of instance methods

• Go test app: Fixed build by adding `-tags kreuzberg_dev` flag for FFI linking
• Go tests: Fixed flawed pointer test that made incorrect assumptions about Go's memory model

• Parity script: Improved `scripts/verify_api_parity.py` to correctly parse all language bindings
  • TypeScript: Better handling of multi-line interfaces with JSDoc
  • Python: Correct parsing of `.pyi` stub files
  • Java: Extract field names from `toMap()` serialization
  • C#: Extract `JsonPropertyName` attributes for canonical names
  • WASM: Dedicated extractor for TypeScript type definitions

• Kubernetes deployment guide: New comprehensive guide for deploying Kreuzberg in Kubernetes (#328)
  • Complete K8s architecture overview with StatefulSet, Service, and ConfigMap examples
  • Health check configuration for plugin readiness and liveness probes
  • Logging aggregation best practices for plugin status visibility
  • Troubleshooting section for silent plugin failures in containerized environments
  • Updated Docker guide with K8s deployment references
  • Location: `docs/guides/kubernetes.md`

Major Release: Complete API Consistency Across All 10 Language Bindings

This release achieves 100% API parity across Rust, Python, TypeScript, Ruby, Java, Go, PHP, C#, Elixir, and WebAssembly bindings. Every ExtractionConfig field is now available in all languages with consistent naming conventions and type safety.

📊 Release Stats:

• 70 files changed
• 11,839 lines added
• 300+ new tests across all bindings
• 4,500+ total tests passing
• Complete backward compatibility for all SDK APIs

• Full config parameter support: All ExtractionConfig options now available via MCP tools
• Enables complete configuration pass-through from AI agents to Rust core
• Standardizes parameter handling across all MCP tools

• --output-format flag: Canonical replacement for --content-format
• --result-format flag: Controls result structure (unified, element_based)
• --config-json flag: Inline JSON configuration
• --config-json-base64 flag: Base64-encoded configuration support

• All 10 bindings now expose identical API surface
• Added missing fields to PHP, Go, and Java bindings
• Fixed Ruby batch chunking operations
• Complete type safety across all languages

• 300+ new tests across all bindings
• API consistency validator with CI integration
• Comprehensive migration guide (757 lines)
• Deprecation guide with code examples

MCP Interface Only (AI agents only, no end-user impact):

• enable_ocr and force_ocr moved under config object
• Old parameter names still work with deprecation warnings

• CLI: --content-format → use --output-format instead
• Environment: KREUZBERG_CONTENT_FORMAT → use KREUZBERG_OUTPUT_FORMAT

All deprecated features remain functional in v4.2.0 and will be removed in v5.0.0.

See the Migration Guide for complete upgrade instructions.

Full Changelog: https://github.com/kreuzberg-dev/kreuzberg/blob/main/CHANGELOG.md#420---2026-01-26

• Ruby: Ruby 4.0 support
  • Updated gemspec to support Ruby 3.2.0 through 4.x
  • Tested with Ruby 4.0.1: all tests pass with Magnus bindings
  • No breaking changes required in binding code

• Ruby: Fixed gem native extension build failure

  • Vendor script now correctly updates native Cargo.toml paths to use vendored crates
  • Fixed sed pattern matching (5 parent directories, not 6)

• Go: Fixed Windows timeout in Go tests

  • Removed init() function in helpers_test.go that caused FFI mutex deadlock on Windows
  • Now uses lazy initialization via sync.Once pattern

• Updated dependencies across the project