Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[4.10.0] - 2026-07-11

First release of the standalone Kreuzberg v4 LTS line. This is the long-term-support home for Kreuzberg v4; active development continues as Xberg (v5+). No behavioural changes to extraction — this release is a security, licensing, and packaging refresh.

Security

Upgraded core dependencies to their latest patched releases (hf-hub, rmcp, lopdf, liter-llm, pyo3, comrak, calamine, quick-xml, text-splitter, tower-http, and others).

Changed

License is now MIT (earlier v4 releases shipped under the Elastic License 2.0).
Repository moved to standalone kreuzberg-dev/kreuzberg-lts. New v4 Go releases publish under github.com/kreuzberg-dev/kreuzberg-lts/v4; existing github.com/kreuzberg-dev/kreuzberg pins keep resolving via the Go module proxy cache.

Added

LTS support policy and migration guide at docs.kreuzberg.dev/lts: v4 receives critical bug and security fixes until the end of 2026 on a best-effort basis. The R binding remains exclusive to the v4 LTS line (not part of Xberg v5).

[4.9.9] - 2026-06-05

LTS patch release with PDF/OCR robustness fixes and selected stability backports from main.

Fixed

OCR on wide, vector-heavy single-page PDFs: uses a bounded render profile and retries at a lower cap instead of failing extraction with PdfiumLibraryInternalError(Unknown).
Embedded PDF image OCR: unsupported image streams are re-extracted through Pdfium or skipped with a specific warning instead of repeated image dimension probe failed noise.
RTF/MSG decompression: cap the initial allocation hint so a crafted stream cannot request a multi-gigabyte allocation from an untrusted size header.
Table row sorting is now NaN-safe (no longer panics).
FFI embedding preset symbols are exported even when embeddings are disabled, preventing Java/native binding startup from failing with missing symbols.
Chunking now uses formatted content for non-plain output formats, preserves page metadata, and normalizes trailing-space page artifacts before boundary matching.
UTF-16 email transcoding: short binary inputs are no longer misclassified as UTF-16.
Returns a validation error when extraction_timeout_secs is set without tokio-runtime.

[4.9.8] - 2026-05-17

LTS patch release. Four targeted bug fixes plus dependency pinning so the branch builds against current crates.io releases.

Fixed

RTF hex escapes now honor \ansicpgNNNN, so CP1251 Cyrillic decodes as readable text instead of Windows-1252 mojibake.
Python ExtractionConfig(cancel_token=…) no longer raises TypeError; the kwarg is accepted and threaded through to the underlying cancellation token.
C# OcrConfig gains the missing VlmConfig property and the previously-undefined LlmConfig type (Model, ApiKey, BaseUrl, TimeoutSecs, MaxRetries, Temperature, MaxTokens).
musl CLI tarball now bundles all transitive ONNX Runtime deps, fixing startup failure on any host.
Build compatibility: pinned tokenizers and v_htmlescape to versions compatible with the current API.

[4.9.7] - 2026-05-08

LTS patch release. Publish-pipeline fixes only — no library code changes.

Maintenance release (internal changes only).

[4.9.6] - 2026-05-07

LTS patch release. Bug fixes backported from main (v5 development). The chore/v4.9-lts branch is now the long-lived line for the 4.9.x series.

Fixed

max_images_per_page cap is now enforced in image extraction itself — previously PDFs with thousands of image objects per page hung indefinitely. Image decoding also moved off the async executor so extraction_timeout_secs can fire while images are processing.
OCR elements are now propagated through the extraction pipeline.
extraction_timeout_secs is now enforced in single-file extraction paths (previously only multi-file batch flows).
PDF image data no longer leaks into structured output when image extraction is disabled.
Preset-only chunking config no longer auto-injects an unwanted EmbeddingConfig on every chunk.
PDF heading and image-placeholder classification corrected.
On the WASM target, OCR PSM now defaults to SINGLE_BLOCK (native default unchanged).
HwpExtractor no longer claims the application/haansofthwpx MIME type (that format is ZIP-based XML, not CFB-based HWP).
Email HTML body fallback: HTML-only emails that the mail parser failed to surface now return content instead of empty output.
Image decode pixel cap: attacker-controlled image bytes above 64 MP are rejected with a clean error instead of triggering multi-GB allocations.
CLI --log-level no longer panics on malformed input (falls back to info).
Markdown rendering: collapse runs of 3+ consecutive newlines to exactly 2.
MCP file_configs schema: emits a schema form accepted by Moonshot AI / Kimi.

API surface

extract_images_from_pdf, extract_images_from_pdf_with_password, and PdfImageExtractor::{extract_images, get_image_count, extract_images_from_page} gain a required max_images_per_page: Option<u32> parameter. Pass None to preserve previous unbounded behaviour.

Not backported from main

The following fixes on main could not be applied to v4.9.x because they build on v5 architecture:

#834 (DOCX inject_placeholders/OCR pipeline integration).
#799 (Form XObject image extraction).
#824 (image extraction across XObject references).

[4.9.5] - 2026-04-23

Fixed

GPU acceleration: kreuzberg now bundles CPU-only ONNX Runtime by default. When a GPU execution provider (cuda, tensorrt, coreml) is explicitly requested but unavailable, it returns an error with setup instructions instead of silently using CPU; Auto mode falls back to CPU with an info log. For GPU support, set ORT_DYLIB_PATH to a GPU-enabled ONNX Runtime.
DOCX OCR extraction: OCR now runs on embedded images and its text is injected into rendered output, instead of being discarded and replaced with placeholder text.
PaddleOCR GPU (CUDA): AccelerationConfig is now propagated to all PaddleOCR ONNX sessions instead of silently falling back to CPU.
PaddleOcrConfig is now exposed in Python bindings (with OcrConfig backward compatibility).
Ruby gem packaging: excludes staged libpdfium.dylib from gem artifacts.

[4.9.2] - 2026-04-19

Fixed

Cancellation token is now checked in the WASM (non-tokio) path for Excel, DOC, PPT, Pages, Keynote, and Numbers extractors — cancellation was previously silently ignored in WASM builds.
Propagate the Cancelled error code (9) to all bindings (Go, C FFI, Python, TypeScript).

[4.9.1] - 2026-04-19

Fixed

Preserve the _internal_bindings.pyi type stub during wheel cleanup — published wheels now include inline type information for the core binding module.

[4.9.0] - 2026-04-18

Fixed

Suppress C23 glibc symbols in manylinux wheels to prevent incompatible symbols on glibc < 2.38 (Debian 12, Ubuntu 22.04).
Remove kreuzberg-cli from the Python wheel to fix libonnxruntime.so.1 loading failure — the CLI is available as a standalone release.
Cancellation token support: cancelled extractions no longer block subsequent calls; wired across Python, Node.js, Ruby, WASM, and C FFI.
Fix kreuzberg[easyocr] extra silently installing nothing on Python 3.14+.
Fix ~1000x slowdown on Ghostscript-produced PDFs with structured output.
Fix llm_usage returning None when using VLM-based OCR.

Added

Cancellation token API available in all language bindings (CancellationToken in Python/Node/Ruby/WASM/FFI).

Changed

Breaking: kreuzberg-cli binary is no longer bundled in the Python wheel — install the standalone CLI from GitHub releases.

[4.9.4] - 2026-04-22

Fixed

Ruby gem build failure — missing max_images_per_page field in ImageExtractionConfig caused a compilation error on all platforms.
PaddleOCR GPU (CUDA): now correctly used when AccelerationConfig(provider="cuda") is set — previously the execution provider was never applied and it silently fell back to CPU.

[4.9.3] - 2026-04-22

Added

Layout detection regions on PageContent — new layout_regions field exposes detected layout regions (class, confidence, bounding box, area fraction) from the RT-DETR model, enabling programmatic detection of diagrams, figures, tables, and other content types per page. Available across all 10 bindings.

Fixed

PaddleOcrConfig in Python API: exposed as a first-class class; OcrConfig accepts both objects and raw dictionaries.
DOCX page extraction (extract_pages=True): result.pages and get_page_count() now work correctly instead of always returning None/0. (Tables spanning multiple pages remain a known limitation.)
serve/mcp CLI subcommands now apply KREUZBERG_* environment overrides (previously only extract honoured them). MISTRAL_API_KEY is now picked up for bare mistral-* model names.
Tagged-PDF structure tree no longer drops paragraph body text when a block has both text and children, and no longer emits malformed markdown for numbered section headings.
Semantic chunker fallback now respects max_characters instead of a hardcoded 4000-char ceiling; warns when chunker_type='semantic' is used without an EmbeddingConfig.
OCR backend dispatch: a non-default backend that errors no longer silently falls back to paddleocr; auto-fallback is limited to the default tesseract backend.
EasyOCR on PDFs: Rust’s renderer now handles page rendering, removing the implicit pdf2image/pymupdf requirement that was never declared in the [easyocr] extra.
OcrConfig.vlm_prompt is now honored in VLM OCR requests (previously documented but never forwarded).
PDF image links are no longer silently dropped from markdown output; respects inject_placeholders.
PDF with large numbers of image fragments no longer hangs — added ImageExtractionConfig.max_images_per_page (default None) and honored extraction_timeout_secs at inter-page checkpoints.
PST extractor now populates email attachments (name, filename, MIME type, size, data); entry IDs formatted as proper MAPI hex strings.

Added

ImageExtractionConfig.max_images_per_page — optional cap on images decoded per page; prevents hangs on PDFs with thousands of inline image fragments.

[4.8.6] - 2026-04-17

Added

PST message EntryID in metadata — the entry_id field from Outlook PST message entries is now included in the metadata of EmailExtractionResult, letting callers link extracted data back to its source message.
AccelerationConfig wired through all ORT model loading — CUDA/CoreML/TensorRT/Auto acceleration is now propagated to all ONNX Runtime sessions (layout detection, embeddings, document orientation, PaddleOCR); previously GPU acceleration was silently ignored. The acceleration field is added to LayoutDetectionConfig and EmbeddingConfig across all 11 bindings.
Semantic chunker (ChunkerType::Semantic) for topic-aware document splitting, with a topic_threshold config field.

Fixed

Batch extraction crash on ARM64 Linux (“Lazy instance has previously been poisoned”) — OCR backend init failures no longer poison a shared static and cascade to all concurrent batch tasks.
PaddleOCR model_tier from TOML ignored by API server — per-request model_tier is now honored.
VLM OCR backend ignored when paddle-ocr feature enabled — vlm_config is now propagated so configured VLM OCR is actually used.
Doubled OCR content and corrupted page text in image extraction — OCR content is no longer duplicated word-by-word into content/pages.
Image OCR pages[] empty — element output is now forced on for image extraction.
LlmConfig missing Default — the documented ..Default::default() pattern now compiles.
LLM embedding provider panics in server mode — embedding no longer panics when called inside an existing tokio runtime (HTTP/MCP server).
OCR table metadata serialized as strings instead of numbers (table_count, tables_detected, table_rows, table_cols), which broke numeric comparisons in all bindings; also fixed a duplicate output_format metadata key.
Ruby structured_output was missing from the Result class and not serialized.
DOCX tables assigned wrong page numbers — now numbered by actual document position.
ocr.enabled=false ignored — OCR no longer runs when explicitly disabled; also fixed a dropped trailing newline in --format text output.

Changed

Updated dependencies including html-to-markdown-rs 3.1→3.2 and tokio 1.51→1.52.

4.8.5 - 2026-04-14

Added

LLM usage tracking — new llm_usage field on ExtractionResult captures token counts, estimated cost (USD), model identifier, and finish reason for every LLM call (VLM OCR, structured extraction, LLM embeddings). Exposed across all bindings.

Fixed

Markdown chunker no longer duplicates a heading when prepend_heading_context is enabled and a chunk boundary aligns with a heading.
Python wheel requires glibc ≥ 2.38 (breaks Debian 12, Ubuntu 22.04) — downgraded to manylinux_2_28 and suppressed C23 glibc symbol emission so wheels install on systems with glibc < 2.38.
FFI memory leak — kreuzberg_free_result now frees djot_content_json, structured_output_json, and llm_usage_json.

4.8.4 - 2026-04-13

Added

Helm chart for Kubernetes deployment — minimal, security-hardened chart (Deployment, Service, Ingress, PVC, HPA, PDB, ServiceAccount), published to GHCR as an OCI artifact.

Fixed

Comrak bridge panic on multi-byte UTF-8 boundaries — annotation offsets landing inside multi-byte characters (e.g. Cyrillic) no longer cause panics.

4.8.3 - 2026-04-12

Fixed

ONNX session creation fails on Linux x86-64 (“graph_optimization_level is not valid”) — the Linux wheel bundled an incompatible ORT version. Switched to a portable optimization level and aligned all ORT versions to 1.24.2.

Documentation

Documented AVX/AVX2 CPU requirement for ONNX Runtime features — CPUs without AVX (e.g. Intel Atom, Celeron N5105) cannot use PaddleOCR, layout detection, or embeddings.

4.8.2 - 2026-04-10

Added

HtmlOutputConfig typed in all bindings — the html_output config field (themes, CSS classes, embed CSS, custom CSS, class prefix) is now fully typed in Python, TypeScript/Node, Go, Ruby, Elixir, PHP, Java, C#, R, and FFI (previously Rust-only).

Fixed

PDF: legitimate repeated content stripped regardless of strip_repeating_text — page-merge deduplication ran unconditionally, removing legitimately repeated brand names and other content even when content filtering was disabled. Both dedup passes are now gated behind the flag.

4.8.1 - 2026-04-09

Added

Styled HTML output — new HtmlOutputConfig on ExtractionConfig with 5 built-in themes (default, github, dark, light, unstyled), semantic kb-* CSS class hooks, CSS custom properties, custom CSS injection, and configurable class prefix. The existing Html output format is upgraded in-place when html_output is set.
5 new CLI flags: --html-theme, --html-css, --html-css-file, --html-class-prefix, --html-no-embed-css — any implicitly sets --content-format html.

Changed

Vendored yake-rust 1.0.3 into core — fixes a BacktrackLimitExceeded panic on large files (10+ MB) and expands YAKE stopwords from 34 to 64 languages.

Fixed

PPTX: panic on non-char-boundary during page boundary recomputation (multi-byte UTF-8 characters).
PDF: include_headers/include_footers flags were ignored by layout-model furniture stripping; they now correctly preserve those regions.
PDF: heuristic table detector no longer misclassifies body text as tables on slide-like PDFs (rejects ≤3-row tables spanning >50% of page height).
PPTX/DOCX/HTML/DocBook/LaTeX/RST: ImageExtractionConfig.inject_placeholders=false now correctly suppresses image references.

4.8.0 - 2026-04-08

Added

Cross-extractor content filtering — new ContentFilterConfig on ExtractionConfig with include_headers, include_footers, strip_repeating_text, and include_watermarks flags across PDF, DOCX, RTF, ODT, HTML, EPUB, and PPT extractors. Typed in all bindings.
Local LLM support via liter-llm 1.2 — use Ollama, LM Studio, vLLM, llama.cpp, LocalAI, or llamafile as VLM OCR, embedding, or structured-extraction backends with zero API key configuration.
LLM-powered document intelligence — integrates with 146 LLM providers for three capabilities: VLM OCR (vision models as OCR backend for low-quality scans, handwriting, Arabic/Farsi, complex layouts, via ocr.backend = "vlm"), structured extraction (JSON-schema-constrained extraction), and VLM embeddings (provider-hosted embedding models).
New CLI command kreuzberg extract-structured, API endpoint POST /extract-structured, and MCP tool extract_structured for schema-guided LLM extraction.
Minijinja template engine for customizable LLM prompts.
5 new environment variables: KREUZBERG_LLM_MODEL, KREUZBERG_LLM_API_KEY, KREUZBERG_LLM_BASE_URL, KREUZBERG_VLM_OCR_MODEL, KREUZBERG_VLM_EMBEDDING_MODEL.
LlmConfig and StructuredExtractionConfig types, structured_output field on ExtractionResult, and EmbeddingModelType::Llm variant across bindings.
Standalone text embedding API with /embed endpoint, embed_text MCP tool, and embed CLI command.

Changed

License changed from MIT to Elastic License 2.0 (ELv2) — copyright holder changed to Kreuzberg, Inc. Forked crates retain their original MIT licenses.
API returns 501 Not Implemented (instead of 500) when the liter-llm feature is disabled.
JSON schema additionalProperties automatically stripped for non-OpenAI providers.

Fixed

PDF: brand names stripped by repeating-text detection — strip_repeating_text = false now disables the removal that incorrectly stripped brand names from PowerPoint-exported decks.
PPTX: slide order scrambled for decks with 10+ slides — fixed lexicographic sort (slide10.xml before slide2.xml) to use numeric ordering.
UTF-8 panic in arXiv watermark stripping when a multi-byte character spans the search limit.
DOC: garbled text from old Word files — CP1252 text misread as UTF-16LE; added heuristic to detect and re-decode.
WASM: table extraction returns empty array when pageNumber is null (now defaults to page 0).

4.7.4 - 2026-04-06

Added

Re-added the --layout boolean CLI flag for easy layout detection enablement (--layout to enable with model defaults, --layout false to disable).
arXiv watermark/sidebar noise filtering for academic PDFs.
Second-tier cross-page repeating-text detection for conference headers and journal running titles outside the margin zone.
Figure/picture text suppression — text inside layout-detected Picture regions is excluded from body output.

Fixed

Figure-internal text leaking into body output — diagram labels and axis text are no longer included (sometimes promoted to headings) in extracted body content.
Empty image references in PDF markdown/HTML output — PDFs with embedded images no longer produce empty ![]() / <img src="">; actual image pixel data is now extracted, producing proper ![](image_N.png) references.
WASM build failure with extern "C-unwind" — added a macro that uses extern "C-unwind" natively and extern "C" on WASM.
Go module tag format — tags now use the correct packages/go/v4/vX.Y.Z format matching the module path.

Changed

CLI documentation updated with missing extraction override flags (--layout-table-model, --disable-ocr, --cache-namespace, --cache-ttl-secs).

4.7.3 - 2026-04-05

Fixed

Archive extraction SIGBUS crash on macOS ARM64 — ZIP, 7Z, TAR, and GZIP extraction crashed in release builds due to miscompilation under opt-level=3; reduced to level 2 for the affected crates.
Native-text PDF extraction fails when OCR backend unavailable — PDFs with extractable native text no longer hard-fail with All OCR pipeline backends failed; the OCR quality-enhancement pass now falls back to the native result with a warning.
Rust cannot catch foreign exceptions crash — C++ exceptions from Tesseract or Leptonica (e.g. on corrupted images) no longer abort the process; they now unwind safely and convert to recoverable errors.

4.7.2 - 2026-04-04

Changed

Global model cache — models now download to a platform-appropriate global cache (~/.cache/kreuzberg/ on Linux, ~/Library/Caches/kreuzberg/ on macOS, %LOCALAPPDATA%/kreuzberg/ on Windows) instead of per-directory .kreuzberg/ folders. Override with KREUZBERG_CACHE_DIR.

Fixed

Embedded HTML in PDF text layers — raw HTML in a PDF text layer (<p>, <br />, <a href>) that previously produced escaped garbage is now converted to clean markdown.
Code classification false positives — regular prose is no longer misclassified as Code blocks.
PageBreak rendering as ----- separators — PageBreak elements no longer pollute output with ----- / <hr>; treated as structural metadata.
Leptonica DPI crash — images with 0 DPI no longer trigger an uncatchable C++ exception; DPI is validated and fixed to 72 before preprocessing.
Node.js ExtractionResult.children missing at runtime in the published v4.7.1 binary.
Node.js disable_ocr config not respected — disableOcr: true no longer produces OCR content for images.
PaddleOCR angle classification crash — fixed input dimensions to match the v2 angle classifier model.
Chunk page numbers missing — chunks produced with first_page/last_page null when chunking was configured without explicit pages config; page tracking is now auto-enabled.

4.7.1 - 2026-04-03

Added

Tree-sitter grammar management CLI — new kreuzberg tree-sitter subcommand with download, list, cache-dir, and clean for managing grammar parsers (by language, group, or all); reads [tree_sitter] config from kreuzberg.toml.
Tree-sitter grammar management API — new REST endpoints POST /grammars/download, GET /grammars/list, GET /grammars/cache, DELETE /grammars/cache.
Tree-sitter grammar management MCP tools — download_grammars, list_grammars, grammar_cache_info, clean_grammar_cache.
Tree-sitter config startup initialization — API and MCP servers auto-download grammars on startup when [tree_sitter] specifies languages or groups.

Changed

Normalized OCR+layout pipeline — the Tesseract+layout path now follows the same architecture as pdfium+layout, fixing destroyed paragraph structure and reading order.
Elixir NIF crash protection — extraction and batch NIFs are wrapped so native-library panics (pdfium, tesseract) return {:error, reason} instead of crashing the BEAM VM.

Fixed

hOCR parser depth tracking — content after inner word spans is no longer silently dropped due to premature paragraph termination.
hOCR multi-page content loss — content on pages 2+ is no longer dropped due to a per-page filter on per-page hOCR documents.
OCR batch parallelization — page processing now scales with available CPUs (capped at 8) instead of a hardcoded 4, speeding up multi-page documents.
Chunking page boundary regression — page boundaries are recomputed from rendered per-page content, fixing null first_page/last_page and a validation warning.
HF Hub environment variables — respects HF_HOME/HF_ENDPOINT, fixing permission errors on Kubernetes when running as non-root.
PDF bridge tracing panic on multibyte characters — no longer panics on multibyte UTF-8 (e.g. •).
Go FFI struct layout — vendored C header was missing children_json, shifting all subsequent fields and reading wrong memory.
Java FFI struct layout — CExtractionResult was missing code_intelligence_json, causing all Java extractions to return success=false.
PHP __get magic method bypass — elements, djotContent, document, ocrElements, children, and uris returned raw JSON strings instead of deserialized arrays.
Ruby disable_ocr config — the keyword was not parsed, so OCR ran even when explicitly disabled.
Node.js ExtractionResult parity — document, djotContent, and ocrElements were omitted from JS objects when None; now default to null.
Node.js convertChunk missing chunkType — the converter did not forward chunk_type.
ODT caption text extraction — text inside draw:text-box (e.g. image captions) was not extracted.
Italian/European PDF ligature corruption — repairs tt, ti, tti ligatures (e.g. Dire*ore → Direttore).
WASM Rayon thread pool panic — parallel iteration now falls back to sequential on wasm32 instead of panicking.

4.7.0 - 2026-03-30

Added

Semantic chunk labeling: chunks now include a chunk_type field identifying the semantic nature of the content (e.g. paragraph, heading, list_item, table_cell, code_block), across all 11 language bindings.
Image extraction across 8 formats: embedded images extracted as ExtractedImage (binary data, format, dimensions, alt text) for DOCX, PPTX, PDF, EPUB, ODT, HTML, RTF, and Markdown/MDX/Jupyter. Markdown output renders as ![alt](image_N.ext) with binary data in ExtractionResult.images.
Recursive OCR on embedded images: extracted images from EPUB, ODT, HTML, and RTF are OCR’d, producing nested ExtractionResult in ExtractedImage.ocr_result.
PDF watermark artifact filtering using pdfium’s /Artifact content marks.
Vertical table header reconstruction: fixes rotated PDF table column headers extracted as reversed spaced characters.
Pages API for PDF extraction: per-page content is now available via result.pages for PDF documents.
TOON wire format: token-efficient (~30-50% fewer tokens) JSON alternative across CLI (--format toon), API (Accept: application/toon), MCP (response_format: "toon"), and all 11 language bindings; losslessly convertible to/from JSON.
Renderer registry: trait-based Renderer/RendererRegistry lets external crates register custom output-format plugins via register_renderer().
URI extraction: new Uri type with UriKind classification (Hyperlink, Image, Anchor, Citation, Reference, Email) extracted from 20+ formats, deduplicated, in ExtractionResult.uris.
Recursive email attachment extraction: EML/MSG/PST attachments and nested message/rfc822 parts extracted as ArchiveEntry children, respecting max_archive_depth.
PDF embedded file extraction: PDF file attachments (portfolios) extracted as children (with filename sanitization, size limits, depth guards).
PDF bookmark/outline extraction: document outlines extracted as URIs (page destinations as anchors, external links as hyperlinks).
DOCX/PPTX embedded object extraction: OLE objects and embedded files extracted as children.
PPTX hyperlink extraction from slide XML, resolved via relationship files.
Image path resolution for markup formats: relative image paths in Markdown, MDX, LaTeX, RST, OrgMode, Typst, Djot, and DocBook are resolved from the filesystem (with path-traversal prevention) and extracted.
FictionBook image and link extraction (base64 <binary> images and <a> hyperlinks).
Apple iWork improvements: Numbers outputs tables, Keynote has improved slide structure, Pages has heading detection; all three extract metadata.
code_intelligence field on ExtractionResult: top-level access to tree-sitter results (structure, imports, exports, chunks, symbols, diagnostics, docstrings).
CodeContentMode config: control code extraction mode — chunks (default), raw, or structure.
TSLP semantic chunking for code: code files use function/class-aware chunks with semantic types and heading context.

Code Intelligence

Tree-sitter integration for 248 programming languages: extract functions, classes, imports, exports, symbols, docstrings, diagnostics; syntax-aware chunking; language detection from extension/shebang; dynamic grammar download (native) / 30-language static subset (WASM); new tree-sitter and tree-sitter-wasm feature flags (in full and wasm-target); TreeSitterConfig/TreeSitterProcessConfig in ExtractionConfig.

Typed Metadata

New FormatMetadata variants: Code, Csv, Bibtex, Citation, FictionBook, Dbf, Jats, Epub, Pst; PptxMetadata extended with image_count/table_count; typed fields across all 11 bindings.

Breaking Changes

Layout detection preset removed: the preset field on LayoutDetectionConfig and the --layout-preset CLI flag are removed; layout detection now uses the RT-DETR v2 model unconditionally. Old configs with "preset" are silently ignored.
Table model config typed: table_model on LayoutDetectionConfig changed from Option<String> to a TableModel enum (tatr, slanet_wired, slanet_wireless, slanet_plus, slanet_auto, disabled), defaulting to tatr. String values still accepted in JSON/TOML.

Fixed

PDF table rendering: populate Table.cells from the TATR/SLANeXT grid so tables render as proper Table nodes.
LaTeX extraction: convert \href, \emph, \textbf, \verb, \sout, blockquotes, lists, special characters, and ligatures to markdown.
XLSX/XLS: emit a ## SheetName heading before each sheet’s table.
OPML/IPYNB/JATS/RST/ORG headings improved (outline headings, notebook ATX headings and cell outputs, JATS abstract/references, RST title levels and code-block hints, ORG source/example blocks).
ODT formula extraction: embedded MathML formulas extracted as formula content instead of empty image placeholders.
PPTX slide titles detected via OOXML placeholder type and emitted as H2 headings; bulleted/numbered lists extracted properly.
RTF formatting: bold/italic/strikethrough use exact byte offsets (fixes formatting bleeding across paragraphs); hidden text suppressed; hyperlink parsing, strikethrough, and multi-row table rendering fixed.
HTML preprocessing: navigation, forms, and sidebars are now stripped by default, removing page chrome from output.
PDF table detection: rejects false tables where >70% of cells are single-word fragments (justified prose).
DocBook root element handling: XML fragments without a root element are wrapped automatically.
PDF image FlateDecode fallback: images that fail decode_flate_to_png() (FlateDecode, CCITT, JBIG2) are re-extracted via pdfium, producing valid PNG instead of unusable bytes.
Metadata standardization: PPTX, Excel, ODT, RST, OrgMode, Typst, RTF, JATS, DOC, PPT, HTML, Email, BibTeX, and Citation metadata now map to standard Metadata fields (title, authors, dates, keywords, language) instead of only the additional map.
MDX/RST/LaTeX/OrgMode/BibTeX/JATS link, URL, and classification fixes (heading and list-item links, RST hyperlinks, \url{} extraction, more image extensions, correct BibTeX URL classification, JATS title field).
Tesseract C++ exception crash: fixed a fatal runtime error where Tesseract C++ exceptions unwound through Rust FFI; Tesseract is now compiled with -fno-exceptions.
ExtractionConfig rejects unknown fields: deny_unknown_fields added, so typos/invalid fields (e.g. layout_analysis) are no longer silently ignored.
PPTX markdown mode derived from output_format instead of hardcoded plain text, so tables/lists render correctly.
DOCX merged cells (gridSpan/vMerge) repeat content across spans; added source_path to ExtractedImage for DOCX image paths.
Python wheel __isoc23_strtoll error on older Linux — downgraded manylinux target to manylinux_2_28 for glibc < 2.39 compatibility (Ubuntu 20.04/22.04, Debian 11/12).
Go macOS link failure — added missing -framework Foundation to CGO LDFLAGS for ORT’s CoreML provider.
Windows GNU ORT linking — uses dynamic linking with pre-downloaded Microsoft ORT; documented the ONNX Runtime DLL requirement for Go, Elixir, and C/C++ on Windows.

Changed

PDF text extraction: full rewrite to page.text().all() + char-indexed font metadata, producing cleaner text with correct word spacing.
CLI format flags: --format (-f) now supports text, json, and toon; --output-format renamed to --content-format (deprecated alias kept with warning).

Removed

max_upload_mb server config field and KREUZBERG_MAX_UPLOAD_SIZE_MB env var — use max_multipart_field_bytes / KREUZBERG_MAX_MULTIPART_FIELD_BYTES instead.
metadata.additional legacy insertions: pipeline features no longer insert error/status keys — errors are in processing_warnings, keywords in extracted_keywords.

4.6.3 - 2026-03-27

Added

Tower service layer (service module): Composable ExtractionService with configurable middleware (tracing, metrics, timeout, concurrency limit) behind a new tower-service feature, auto-enabled by api/mcp.
Semantic OpenTelemetry conventions: Formal kreuzberg.* attribute namespace with span attributes, metric names, and operation/stage constants for extraction, pipeline, OCR, and model inference telemetry.
Extraction metrics: 11 OTel instruments covering extraction totals, durations, cache hits/misses, pipeline stages, OCR, and concurrency (feature-gated behind otel).

Improved

Deeper instrumentation: Pipeline post-processing stages, individual processors, OCR, and layout model inference now emit semantic spans and duration metrics.
API and MCP servers route extractions through the Tower service stack, gaining unified tracing, metrics, and middleware.
API server hardening: Added response compression (gzip/brotli/zstd), panic recovery, request-ID correlation, and sensitive-header redaction.

Changed

Span attribute names migrated to kreuzberg.* namespace (e.g. extraction.filename -> kreuzberg.document.filename).

Fixed

EPUB spine handling: Preserves manifest fallback chains, guide references, and non-linear spine items; strips navigation chrome; malformed guide references now warn instead of hard-failing.
DOCX images with high-quality settings not extracted: <a:blip> elements with child elements are now handled.
OCR tables discarded on default output format: Layout detection was skipped for the Plain format, dropping tables from the OCR pipeline; both paths now propagate tables.
Missing binding fields: Exposed chunker_type, sizing_cache_dir, and prepend_heading_context across Python, TypeScript/WASM, Go, C#, and PHP.
Full API parity across all 10 bindings: Added max_archive_depth everywhere, acceleration/email to Ruby/R, layout to PHP, and 7 missing fields to WASM.
Windows install failure for the Node package: Replaced bash-specific prepare-script fallback with a cross-platform equivalent.

4.6.2 - 2026-03-26

Added

PDF page rendering API: New render_pdf_page function and PdfPageIterator render individual PDF pages as PNG images across all 11 bindings, with idiomatic patterns per language. Default 150 DPI, configurable per call.

Fixed

Zero tables found on scanned PDFs: Layout bboxes (640x640 model space) are now scaled to OCR render resolution before table recognition.
OCR elements reported page_number: 1 for all pages: Page numbers are now correctly stamped after OCR in the batch loop.
Ruby gem missing ONNX Runtime: Added bundled ORT so OCR/layout/embeddings work out of the box.

Improved

version:sync also syncs Go C header, version constant, and Docker tags, preventing version drift across bindings.

4.6.1 - 2026-03-25

Added

Per-file batch extraction timeouts: New extraction_timeout_secs (batch default) and timeout_secs (per-file override), plus a KreuzbergError::Timeout variant with elapsed/limit fields. All bindings updated.
Page-level OCR overrides: New force_ocr_pages (1-indexed) enables selective OCR on specific pages of mixed-quality PDFs while preserving native text on others.
PST extraction support: Extract emails from Microsoft Outlook PST archives, feature-gated under email.
JSONL/NDJSON extraction: Native .jsonl/.ndjson extraction registered as application/x-ndjson.

Fixed

OCR elements now propagated to ExtractionResult with geometry data and coordinate-bearing narrative blocks.
OOM crash on multi-page scanned PDFs: PDF pages are now rendered and OCR’d in bounded batches instead of all at once.
OCR memory usage reduced 60-78%: One page is rendered and encoded at a time; a 98-page scanned PDF dropped from 4.6GB to 1.9GB peak RSS. Batch size adapts to available memory on Linux/macOS.
PDF control-character artifacts: Broken ToUnicode mappings that produced control characters where hyphens belong are fixed (e.g. re\x02labelling -> re-labelling).
Missing headings for PDFs: Heading nodes are now inserted for structured PDFs, and markdown heading markers in fallback paragraphs create heading groups.
Empty tables on scanned PDFs: Three bugs that always returned [] tables for scanned/image PDFs are fixed so all paths propagate tables.
Table recognition coordinate mismatch on scanned PDFs: Bounding boxes are now scaled from layout-model resolution to actual OCR render resolution, fixing zero recognized tables.
OCR elements reported page_number: 1 for all pages: Correct 1-indexed page numbers are now stamped per batch page.
PDF layout engine crash on malformed input: Layout-engine init failure now returns a descriptive error instead of crashing the host process via FFI.

4.6.0 - 2026-03-24

Added

Recursive archive extraction: ZIP/TAR/7Z/GZIP archives now recursively extract every processable file, each with its own ExtractionResult. New ArchiveEntry type; configurable max_archive_depth (default 3, 0 for legacy single-text behavior).
YAML/JSON section chunker: New ChunkerType::Yaml splits structured files by keys with full hierarchy paths, auto-inferred for YAML/JSON.
Unified DocumentStructure DTO: 7 new node types, 4 new annotation kinds, and a format-specific attributes bag on every node.
Unified rendering module: render_to_markdown() and render_to_plain() walk a DocumentStructure tree for consistent output with inline annotations, table escaping, and nested lists.
DocumentStructure support for all 35 formats: Every extractor natively produces a DocumentStructure when include_document_structure is enabled, including Office, markup, books, scientific, data, email, and image formats.
DocBook/JATS inline annotations: Semantic emphasis, bold, code, links, and sub/superscript for academic documents.
Document-level OCR: OcrBackend supports whole-file process_document(), up to 30% faster on multi-page documents.

Changed

CSV extraction for embedding quality: Emits Row N: Header: Value format when a header row is detected; the tables field is unchanged.
XML extraction for embedding quality: Indented hierarchical output preserving the element tree with inline attributes and xmlns:* filtering.

Improved

Zero-copy file I/O: Automatic memory-mapping for files >1MB with SIMD-accelerated UTF-8 validation; measurable speedup for large PDFs and archives (WASM falls back to heap allocation).
Unified concurrency management: Centralized thread budget for Rayon, ONNX, and PaddleOCR via configurable ConcurrencyConfig; PDF OCR batched to reduce memory on large documents.

Fixed

Incorrect page numbers in element-based output: Element-based output now auto-enables page extraction, fixing all elements reporting page_number=1.
MSG extraction misses compressed RTF bodies: Added compressed-RTF fallback so .msg files that store the body only in compressed RTF are extracted.
Indexed-colour PDF images returned as raw bytes: Palette-based PDF images now decode to valid PNG output.
ODT extraction robustness: Replaced unwraps with safe fallbacks in ODT parsing.

4.5.4 - 2026-03-23

Added

Document-level OCR optimization: OcrBackend supports native whole-file process_document(), avoiding per-page PDF rasterization when the backend supports it.
PST (Outlook Personal Folders) extraction: New PstExtractor extracts subject, sender, recipients, body, and date from every message, enabled via the email feature (MIME application/vnd.ms-outlook-pst).
prepend_heading_context chunking option: When set with the Markdown chunker, prepends the heading hierarchy path to each chunk for self-contained RAG context. Available across all 10 bindings, CLI, and WASM.

Fixed

PDF image extraction panic on mismatched buffer lengths: Malformed PDF images are now skipped instead of panicking (regression from 4.5.0).
pdf feature failed to compile without layout-detection: Feature-gated the offending config.layout reference.
Ruby binding missing table_model field: Added table_model parsing to the layout config.
WASM fails to load in Supabase/Deno edge functions: Added explicit package exports and Deno detection with clear error messaging for restricted edge runtimes.
Build failures on some toolchains and WASM/CI: Pinned zip below 7.4 and vendored HWP text extraction to drop a transitive zip 2.x dependency.

4.5.3 - 2026-03-22

Added

Apple iWork format support: Native parsing of modern .pages, .numbers, and .key files via a new iwork feature.
SLANeXT table structure recognition models: New table_model field on LayoutDetectionConfig selects the backend (tatr default, slanet_wired, slanet_wireless, slanet_plus, slanet_auto). Available across all 12 bindings and CLI (--layout-table-model).
PP-LCNet table classifier: Automatic wired/wireless table detection for SLANeXT auto mode.
CLI cache warm --all-table-models: Opt-in download of SLANeXT model variants (~730MB); default warm downloads only RT-DETR + TATR.

4.5.2 - 2026-03-21

Fixed

PDF word splitting in extracted text: Spurious mid-word spaces (e.g. "s hall a b e active") are fixed via selective page-level respacing using character-level gap analysis.
Markdown underscore escaping removed: Identifiers like CTC_ARP_01 are no longer mangled into CTC\_ARP\_01.
Page header/footer leakage: Running headers and copyright footers are now detected via fuzzy alphanumeric matching and stripped from the body.
R batch functions failed with a spurious NULL argument: Removed the extra positional NULL that broke all batch operations.
Elixir OCR/layout/embeddings failed on Windows: ONNX Runtime DLL is now staged where the BEAM VM loads NIFs.

Added

General extraction result caching: All file types (not just OCR) are now cached; repeat extractions with the same config return instantly.
Cache namespace isolation: New cache_namespace (and --cache-namespace) enables multi-tenant cache isolation on shared filesystems, with per-namespace deletion and stats.
Per-request cache TTL: New cache_ttl_secs overrides the global TTL per extraction (0 skips cache).
Bundled eng.traineddata: English OCR works out of the box with zero runtime configuration (~4MB bundled).
Tessdata in cache warm/cache manifest: cache warm downloads all ~120 tessdata_fast language files, giving full Tesseract language support without system packages; KREUZBERG_CACHE_DIR/tessdata is now resolved.
CLI embed command: Generate vector embeddings from text (feature-gated on embeddings).
CLI chunk command: Split text into chunks with configurable size, overlap, chunker type, and tokenizer.
CLI completions command: Generate shell completions for bash, zsh, fish, and powershell.
CLI --log-level global flag: Override RUST_LOG per invocation.
CLI extraction overrides: 27 new flags including --layout-preset, --acceleration, --extract-images, --target-dpi, --token-reduction, --max-concurrent, and more.
CLI colored output: Colored headers and labels, respecting NO_COLOR.
API POST /detect, GET /version, GET /cache/manifest, POST /cache/warm: New MIME detection, version, model manifest, and eager-download endpoints.
MCP get_version, cache_manifest, cache_warm, embed_text, chunk_text tools.
TATR model availability check: Layout detection now errors when table regions are detected but the TATR model is unavailable, instead of silently degrading.

Changed

CLI batch flags: The batch command now supports all extraction override flags, matching the extract command.

Improved

CLI/API/MCP input validation: OCR backend names, chunk sizes/overlap, DPI range, layout confidence, embedding preset names, and chunk bounds are now validated.
Chunk overlap auto-clamping: Overlap is clamped to size/4 when --chunk-size is smaller than the default overlap, instead of erroring.

4.5.1 - 2026-03-20

Java batch extraction failed with memory access errors: Fixed swapped count/results fields in the CBatchResult Panama FFM layout.
Go bindings read fields at wrong offsets: Synced the Go C header struct field order with the Rust #[repr(C)] layout, fixing corrupted pages_json.
FFI failed to compile without layout-detection: Feature-gated LayoutDetectionConfig and the layout setter.
Python wheel builds failed on Linux aarch64: The OpenSSL path is now detected via uname -m instead of being hardcoded to x86_64.
R batch functions errored: Added the missing file_configs parameter that broke all batch operations.
R package failed to load with ONNX Runtime: Now links against ORT when ORT_LIB_LOCATION is set, fixing undefined symbol: OrtGetApiBase.

4.5.0 - 2026-03-20

Added

ONNX-based document layout detection: New layout config field runs RT-DETR v2 (17 element classes) with "fast"/"accurate" presets and auto-downloaded models, across all bindings.
SLANet table structure recognition: Detected table regions produce markdown tables with colspan/rowspan support, now on all pages.
Layout-enhanced heading detection: Layout SectionHeader/Title regions guide heading detection and can override font-size classification.
Multi-backend OCR pipeline: New OcrPipelineConfig enables quality-based fallback across OCR backends (e.g. Tesseract then PaddleOCR).
OCR quality thresholds: New OcrQualityThresholds with 16 tunable parameters for output quality assessment and fallback.
OCR auto-rotate: New OcrConfig.auto_rotate (default false) detects 0/90/180/270-degree page rotations.
PaddleOCR v2 model tier system: New model_tier field with "mobile" (default, fast) and "server" (highest accuracy), unified multilingual models, across all bindings.
AccelerationConfig for GPU/execution-provider control: Fine-grained control over ONNX providers (CPU, CoreML, CUDA, TensorRT), typed across all bindings.
ConcurrencyConfig for thread limiting: New max_threads caps Rayon, ONNX intra-op threads, and batch concurrency, typed across all bindings.
EmailConfig for MSG fallback codepage: Configurable fallback codepage for MSG files lacking one (default windows-1252), typed across all bindings.
Per-file extraction configuration (FileExtractionConfig): Each file in a batch can specify its own OCR, chunking, and output settings, via CLI --file-configs and MCP file_configs.
Opt-in single-column pseudo tables: New allow_single_column_tables on PdfConfig (default false) emits glossaries and itemized lists as tables.
CLI cache warm and cache manifest commands: Eagerly download all OCR/layout models, or output a JSON manifest with SHA256 checksums and source URLs.
ChunkSizing configuration: sizing_type, sizing_model, and sizing_cache_dir exposed across all bindings.
Chunk heading context: New HeadingContext in ChunkMetadata with heading level and text.
ModelManifestEntry type with manifest() / ensure_all_models(): Public API for querying and eagerly downloading model caches.

Changed

Layout preset default changed from "fast" to "accurate": The Fast variant is removed; the "fast" string is still accepted.
PaddleOCR default model tier changed from "server" to "mobile": Mobile is 3-5x faster with equivalent quality on standard documents; server remains available.
PaddleOCR v2 models: All models updated to v2 generation with unified multilingual recognition; V1 models remain for older versions.
Batch API unification: _with_configs batch functions removed; per-file FileExtractionConfig is now an optional parameter on the unified batch functions.
Embedding embed() now takes &self: Enables parallel embedding generation.
padding field in PaddleOcrConfig: Now exposed across Python, TypeScript, Ruby, and Go.
Language-agnostic section pattern recognition: Headings ending with a period are allowed when matching structural patterns, improving heading detection for legal/academic/multilingual documents.
Strong typing across bindings: Replaced weak Dictionary/Map/array config types with strongly typed classes in C#, Java, and PHP, and added missing config types to other bindings.

Removed

fastembed dependency: Replaced by a vendored embedding engine using ONNX Runtime directly.
EmbeddingModelType::FastEmbed variant: Use Preset or Custom instead.

Fixed

C# library completely broken in 4.4.6: Fixed the CExtractionResult struct layout mismatch that caused deserialization failures and overflow exceptions.
PDF force_ocr without explicit OCR config silently ignored: Now unconditionally triggers OCR with default settings.
PDF image extraction returned raw compressed data: Extracted images are now decoded and re-encoded as PNG/JPEG.
Node.js extractFileInWorker dropped mime_type: MIME type is now forwarded to extraction instead of injected into the PDF password config.
DOCX/XML parser compilation failure: Resolved type ambiguity introduced by zip 8.2.0.
Python type stubs missing from sdist: py.typed and .pyi files are now included in wheel and sdist.
PDF broken CMap word spacing: Geometric validation vetoes false word boundaries (e.g. "co mputer" -> "computer").
PDF structure-tree headings rejected: H1-H6 tags are now trusted as author intent instead of being rejected by font-size validation.
Slow PDF structure-tree extraction: Text and style maps are built in a single pass, eliminating multi-second extraction on complex pages.
OCR Picture regions suppressed text: Embedded text in Picture regions is preserved as paragraphs instead of dropped.
Unstable reading order: Spatial sorts use discrete row buckets, ensuring correct and stable ordering.
Page furniture over-stripping: Added guards to prevent removal of legitimate content.
KREUZBERG_CACHE_DIR not respected by all caches: Embeddings, OCR, and extraction caches now honor the variable.
MSG ANSI string decoding: MSG files now decode ANSI properties using the declared code page instead of lossy UTF-8.
SLANet-Plus inference failures on macOS CoreML: Re-exported the ONNX model with a shape fix, resolving silent table-extraction failures.
TATR model crash in batch processing: Model unavailability in parallel closures now falls back gracefully instead of crashing FFI callers.
Docker musl builds failed: Alpine/musl images now link the system ONNX Runtime; all features work in musl CLI images.
C#/Java FFI batch functions rejected NULL: They now accept NULL for per-file config JSON.

Known Issues

PHP PIE Windows package temporarily unavailable: Disabled due to a transitive dependency conflict on the Windows target; Linux and macOS are unaffected.
WASM: no layout detection, acceleration, or email config: ONNX Runtime is unavailable on WASM, so RT-DETR layout detection, hardware acceleration, and concurrency config are unavailable; Tesseract WASM OCR and embeddings are supported.

4.4.6

Added

dBASE (.dbf) format support: Extract table data from dBASE files as markdown tables with field type support.
Hangul Word Processor (.hwp/.hwpx) support: Extract text from HWP 5.0 documents.
Office template/macro format variants: Added .docm, .dotx, .dotm, .dot, .potx, .potm, .pot, .xltx, and .xlt support.

Fixed

DOCX image placeholders missing with extract_images=True: The default plain-text path stripped image references; image extraction now forces markdown output so placeholders appear.

Changed

Format count updated to 91+ (from 75+) across docs and manifests.

4.4.5

Fixed

PDF markdown garbled positioned text: Line breaks in positioned/tabular text (CVs, addresses, tables) are preserved by splitting short lines into separate paragraphs when few lines reach the right margin.
Node worker pool password bug: extractFileInWorker passed password as mime_type, so passwords were never applied; it now injects into config.pdf_options.passwords.
WASM document structure returned empty results: JS camelCase config keys are now transformed to snake_case for deserialization.
PHP 8.5 array coercion on macOS: A wrapper transparently converts coerced array return values back to objects.
PHP 8.5 support: Upgraded ext-php-rs to 0.15.6.

Added

CLI --pdf-password flag: New option on extract/batch for encrypted PDFs (repeatable).
MCP pdf_password parameter: Added to extract_file, extract_bytes, and batch_extract_files.
API pdf_password multipart field: The extract endpoint now accepts encrypted-PDF passwords.

4.4.4

Maintenance release (internal changes only).

4.4.3

Added

PDF image placeholder toggle: New inject_placeholders on ImageExtractionConfig (default true); set false to extract image data without injecting ![image](...) references.

Fixed

Token reduction not applied: Config was accepted but never executed; the pipeline now applies reduce_tokens() when configured.
Nested HTML table extraction: Nested tables now extract correctly with proper cell data and markdown rendering.
hOCR plain text output: hOCR now produces plain text when requested instead of falling back to Markdown.
PDF garbled positioned/tabular text: Spaces are now inserted when the X-position gap between characters exceeds 0.8 × avg_font_size.
Chunk page metadata drift with overlap: Byte offsets are computed via pointer arithmetic, fixing incorrect page numbers when overlap is enabled.
Node.js metadata casing: Standardized Metadata/EmailMetadata fields to camelCase (e.g. pageCount, creationDate) and corrected authors/keywords pluralization.

4.4.2

Fixed

WASM OCR blocked the event loop: ocrRecognize() ran synchronously on the main thread; it now runs in a worker thread, keeping the runtime responsive and letting timeouts fire.
JPEG 2000 / JBIG2 OCR decode failure: JP2/JPX/JPM/MJ2 and JBIG2 images failed with “format could not be determined”; a shared loader now detects them by magic bytes and uses the correct decoders across all OCR backends, including ocr-wasm.
WASM PDF returned empty content: initWasm() now awaits PDFium initialization, fixing a race where extraction started before PDFium was ready.

Added

OMML-to-LaTeX math conversion for DOCX: DOCX equations are now converted to LaTeX (fractions, radicals, n-ary operators, matrices, etc.) with $$...$$/ $...$ in markdown and raw LaTeX in plain text, instead of concatenated Unicode.
Plain text output paths for all extractors: DOCX, PPTX, ODT, FB2, DocBook, RTF, and Jupyter now produce clean plain text (no #, **, |, image placeholders) when plain/structured output is requested, instead of always emitting markdown.

Changed

CLI now built with the full feature set (the cli feature group is removed), ensuring the CLI supports all formats including archives (7z, tar, gz, zip).

Fixed

Alpine/musl CLI Docker image “Dynamic loading not supported”: The CLI binary is now dynamically linked against musl libc, enabling runtime loading for PDF processing.
R package Windows installation: Improved Python detection (py launcher, RETICULATE_PYTHON) and graceful handling of symlink extraction errors.
PHP 8.5 precompiled binaries: Added PHP 8.5 support alongside PHP 8.4.
OCR DPI normalization: Images are normalized to the target DPI before Tesseract, eliminating the resolution-estimate warning and improving accuracy on non-standard DPI.
HTML metadata missing with Plain output: HTML headers, links, images, and structured data are now collected even for the default plain output.
PPTX text run spacing: Adjacent runs are joined with smart spacing (“HelloWorld” -> “Hello World”).
CSV Shift-JIS/cp932 detection: Shift-JIS CSVs are decoded correctly instead of producing mojibake, with fallback encoding detection.
EML multipart body extraction: All text/html parts and nested message/rfc822 parts are now extracted, not just index 0.
EPUB media tag leakage: <video>, <audio>, <iframe>, etc. no longer leak into text; <br>/<hr> become newlines.
FB2 poem extraction: <poem>, <stanza>, and <v> verse elements are now extracted instead of dropped; added sub/superscript, strikethrough, and footnotes.
ODT StarMath-to-Unicode conversion: ODT formulas are converted to Unicode instead of raw StarMath syntax.
BibTeX output format: Now uses standard @type{key, field = {value}} conventions.
LaTeX display math: \[...\] environments are converted to $...$ .
RST directive preservation: Field lists, directive markers, and .. code-block:: directives are preserved.
Typst extraction improvements: Layout directives stripped, headings as plain text, column-aware tables, links as display text.
DOCX field codes refined: Field results (visible text like “Figure 1:”, page numbers) are preserved while field instructions are skipped.
DOCX drawing alt text in plain text: Image alt text from wp:docPr is now emitted instead of skipped.
DOCX XML entity decoding: Attribute values like 
 are now correctly unescaped.

4.4.1

Added

OCR table inlining into markdown content: With Markdown output, OCR-detected pipe tables are inlined into result.content at their correct vertical positions instead of only appearing in result.tables.
OCR table bounding boxes: OCR-detected tables now include pixel-level bounding boxes, propagated through all bindings as Table.bounding_box.

Fixed

MSG recipients missing email addresses: The extractor now reads recipient substorages for full "Name" <email> output with correct To/CC/BCC separation.
MSG date missing or incorrect: Date is now read from PR_CLIENT_SUBMIT_TIME, with fallback to transport headers.
EML date mangled for ISO 8601 formats: The raw Date: header text is preserved instead of being garbled.
EML/MSG attachments line polluted text output: The Attachments: ... line is removed from text output; names remain in metadata.
HTML script/style tags leaked in email fallback: Multiline <script>/<style> content is now stripped (dotall matching).
SVG CDATA leaked JavaScript/CSS: <script>/<style> CDATA no longer appears in SVG text output.
RTF metadata noise in text: Destination groups (font tables, color tables, stylesheets, info) are now skipped, removing ~17KB of internal noise.
RTF \u control words mishandled: Formatting commands like \ul are no longer misread as Unicode escapes.
RTF paragraph breaks collapsed to spaces: \par now emits newlines, and whitespace normalization preserves paragraph structure.

4.4.0

Added

R language bindings: New kreuzberg R package via extendr with the full extraction API (sync/async, batch, bytes), typed errors, S3 result class, config discovery, and OCR/chunking configuration.
PHP async extraction: Non-blocking extractFileAsync(), extractBytesAsync(), and batch variants via a Tokio thread pool, with Amp and ReactPHP bridges.
WASM native OCR (ocr-wasm): Tesseract compiled into the WASM binary, enabling OCR in Browser/Node/Deno/Bun without browser-specific APIs; 43 languages via CDN tessdata.
WASM Node.js/Deno PDFium support: PDFium now loads from the filesystem, configurable via KREUZBERG_PDFIUM_PATH.
WASM full-feature build: OCR, Excel, and archive extraction are enabled by default in the WASM package.
WASM Excel extraction (excel-wasm): Calamine-based spreadsheet extraction without a Tokio runtime.
WASM archive extraction: ZIP, TAR, 7z, and GZIP extraction now available in WASM.
WASM PDF annotations: PDF annotations are exposed via the annotations field on ExtractionResult.
C FFI distribution: Official libkreuzberg shared library with cbindgen header, cmake/pkg-config packaging, and prebuilt binaries for Linux x86_64/aarch64, macOS arm64, and Windows x86_64.
Go FFI bindings: New Go package (packages/go/v4) consuming the C FFI library, with prebuilt binaries for all four platforms.
R distribution via r-universe: Switched from CRAN to r-universe for faster releases and easier native compilation.

Fixed

DOCX equations not extracted: OMML math (<m:oMath> etc.) was silently dropped; math runs are now extracted as text.
DOCX line breaks ignored: <w:br/> now inserts whitespace instead of merging adjacent text.
PPTX/PPSX table content lost: Tables now render as markdown pipe tables with proper cell separation instead of an unreadable blob.
PPTX/PPSX/PPTM and DOCX image markers polluted text: Image references now use a clean ![image]() / ![alt](image) form instead of injecting numeric tokens.
EPUB double-lossy conversion: XHTML is now traversed directly instead of through an XHTML->markdown->plain-text pipeline that lost content.
Excel float formatting dropped numeric precision: Whole-number floats format as "1" instead of "1.0".
HTML metadata polluted content: YAML frontmatter is no longer prepended to content when metadata is already returned as a struct.
Markdown extractor lost tokens through AST reconstruction: Now returns raw text content directly while still parsing the AST for tables and images.
SVG text extraction included element prefixes: SVG extraction now targets only text-bearing elements without element_name: prefixes.
WASM OCR not working (enableOcr() regression): enableOcr() now registers the backend in the Rust-side registry so OCR works via extractBytes/extractFile.
WASM tessdata CDN URL returned 404: Updated to the official tesseract-ocr/tessdata_fast repository.
XML UTF-16 parsing failed on odd byte counts: The decoder now truncates to the nearest even byte boundary instead of rejecting the file.
R bindings crashed on strings with embedded NUL bytes: NUL bytes are now stripped before passing strings to R.
R bindings %||% operator incompatible with R < 4.4: Added a package-local polyfill.
API returned HTTP 500 for unsupported formats: The /extract endpoint now falls back to extension-based MIME detection and maps UnsupportedFormat to HTTP 400.
PDF markdown missing headings/bold for flat structure trees: PDFs tagging everything as <P> (e.g. InDesign) now get headings via font-size clustering, and bold detection recognizes “Bold” in font names.
PaddleOCR backend not found via backend="paddleocr": The registry now resolves the "paddleocr" alias to the canonical "paddle-ocr" name.
WASM metadata serialization dropped fields: Switched to serde_json + JSON.parse() so format_type and format-specific metadata are preserved.
WASM config deserialization: camelCase config keys are converted to snake_case before crossing the WASM boundary.
WASM PDFium module loading: The build now copies the real PDFium ESM module instead of a stub, with a Deno compatibility fix.
Email header extraction lost display names: From/To/CC/BCC now use "Display Name" <email> format when available.
Email date header normalized to RFC 3339: The raw Date header value is now preserved, falling back to RFC 3339 only when unavailable.

Removed

polars dependency: Removed unused crate and dead code; Excel extraction uses calamine directly.

4.3.8

Added

MDX format support (mdx feature): extract text from .mdx files, stripping JSX/import/export syntax while preserving markdown, frontmatter, tables, and code fences.
List supported formats API: query all supported extensions and MIME types via list_supported_formats(), the GET /formats REST endpoint, the list_formats MCP tool, or kreuzberg formats.

Fixed

PDF ligature corruption in CM/Type1 fonts: repairs broken-CMap text like di!erent → different, #nancial → financial in LaTeX-generated PDFs.
PDF dehyphenation across line boundaries: rejoins words broken across line breaks (e.g. soft ware → software), including implicit breaks where the hyphen was stripped.
PDF page markers missing in Markdown and OCR output: insert_page_markers / marker_format were dropped for Markdown and OCR output since 4.3.5; markers now appear (Djot inherits them).
PDF Djot/HTML output quality parity: Djot and HTML now use the same structural pipeline as Markdown (headings, tables, bold/italic, dehyphenation) instead of falling back to plain paragraphs.
PDF sidebar text pollution: widened the margin band so rotated sidebar text (e.g. arXiv identifiers) no longer leaks into extracted content.
Node.js PDF config dropped: extractAnnotations, hierarchy, topMarginFraction, and bottomMarginFraction were silently discarded, making PDF annotation extraction always return undefined.

4.3.7

Added

NFC unicode normalization applied to all extraction outputs for consistent composed characters across backends (gated behind quality feature).
Configurable PDF page margin fractions (top_margin_fraction, bottom_margin_fraction) in PdfConfig.
PDF annotation extraction via extract_annotations in PdfConfig and a new annotations field on ExtractionResult (all bindings), with a PdfAnnotation type covering Text, Highlight, Link, Stamp, Underline, StrikeOut, and Other.

Fixed

PDF markdown extraction quality at parity with docling (91.0% vs 91.4% avg F1, 10-50x faster): switched to per-character extraction that correctly handles font matrices, CMap lookups, and positioning, with adaptive line-break detection.
PDF markdown no longer drops all content on broken font metrics: falls back to unfiltered segments when the min-font-size filter would remove everything.
PDF markdown no longer drops all content on edge-case margins: skips margin filtering for a page when it would remove all text.
PDF multi-column extraction improved (69.9% → 90.7% F1 on Federal Register-style layouts) via correct reading order.
PDF table detection now requires ≥3 aligned columns and rejects low-quality tables, eliminating false positives from two-column text layouts.
PDF markdown no longer drops titles/authors when pdfium returns zero-value baseline coordinates (some LaTeX PDFs).
PaddleOCR backend validation now checks the plugin registry, preventing false “backend not registered” errors when the plugin is available.
WASM bindings now export detectMimeFromBytes and getExtensionsForMime.
XLSX extraction with output_format="markdown" now produces markdown tables instead of plain text.
MCP no-parameter tools (cache_stats, cache_clear) now emit a valid object inputSchema, fixing MCP clients that validate schema type.
Python get_valid_ocr_backends() now includes paddleocr, matching other bindings.
PHP PdfConfig now includes extractAnnotations, topMarginFraction, and bottomMarginFraction, restoring parity with the Rust core.

4.3.6

Added

Automatic OCR fallback on garbled fonts: pages where >30% of characters have broken unicode mappings (tofu) now trigger OCR automatically.
Extended list detection: recognizes en/em dashes, single-letter prefixes (a., B)), and roman numerals.
Soft-hyphen and stray control-character normalization in PDF text for cleaner word rejoining.

Fixed

PDF list detection panic on multi-byte UTF-8 with CRLF line endings: fixed crash from byte-vs-char newline handling.
PaddleOCR backend not respected in Python bindings: paddleocr/paddle-ocr backends now correctly hand off to the Rust core instead of being silently skipped.
Ruby gem missing sorbet-runtime at runtime: promoted from a dev to a runtime dependency, fixing load failures.
DOCX extractor panic on multi-byte UTF-8 page boundaries: fixed crash from byte-index slicing during page-break insertion.
Node.js djot_content always undefined: the field is now mapped from Rust results.
Tesseract OCR word-level elements dropped: fixed off-by-one TSV level mapping and a hardcoded None in the image OCR path so word-level elements are returned.
OCR cache deserialization failure: older cached OCR data missing the newer detection field now deserializes again.
PDF table detection false positives: precision improved from 50% to 100% by validating tables on both the pdfium and OCR paths.
PDF line grouping and paragraph-break detection made robust to subscripts/superscripts, fixing mis-grouped lines and spurious paragraph breaks.

4.3.5

Added

PDF markdown output format: native PDF extraction now supports output_format: Markdown with headings, paragraphs, inline bold/italic, and list detection instead of flat text.
Multi-column PDF layout detection: 2+ column layouts (academic papers, magazines) are processed per column, preventing text interleaving.
musl/Alpine Linux prebuilt native libraries for Elixir, Java, and C#: instant install on Alpine and musl distros without compiling from source.
Pre-compiled platform-specific Ruby gems (x86_64-linux, aarch64-linux, arm64-darwin, x64-mingw-ucrt): eliminates the 30+ minute compile on gem install kreuzberg (source gem still available as fallback).
bounding_box field on Table and ExtractedImage (all bindings): spatial positioning for tables and images, computed from character positions.
Inline tables and image placeholders in PDF markdown: tables and image references (with OCR text as blockquotes) are now placed at their correct vertical position instead of appended at the end.
Typed ExtractionResult fields: extracted_keywords (with algorithm/score/position) and quality_score are now first-class typed fields instead of untyped metadata.additional entries (all bindings).
ProcessingWarning on ExtractionResult.processing_warnings: surfaces non-fatal warnings during extraction.
Typed Metadata fields: category, tags, document_version, abstract_text, and output_format; output_format is now populated for every output format, not just structured.

Fixed

PaddleOCR wrong-character recognition: fixed an off-by-one dictionary offset that mismapped recognized characters.
PaddleOCR angle classifier misfiring on short text: use_angle_cls now defaults to false (re-enable via PaddleOcrConfig::with_angle_cls(true) for rotated docs).
PaddleOCR padding included table gridlines: default detection padding reduced 50px → 10px and made configurable via with_padding().
PaddleOCR recognition height mismatch: fixed [batch, 3, 48, width] input shape that caused ONNX Runtime dimension errors on all platforms.
Faster PDF markdown: eliminated a redundant document parse, saving 25-40ms per PDF.
PDF crashes on corrupt coordinates: replaced panics in text clustering/merging with graceful handling of NaN/Inf coordinates.
PDF heading and list false positives: decorative extreme font sizes no longer become headings; long paragraphs starting with “1.” or “-” are no longer misclassified as lists.
PHP binding: bounding_box always null on Table and ExtractedImage.
Go binding: embedding config dropped: added the missing Embedding field to ChunkingConfig, so embedding-enabled extractions actually run embeddings.
Go binding: extracted_keywords, quality_score, processing_warnings always nil: the FFI header and decoder now include these fields.
extraction_duration_ms missing from Go, Java, PHP, C# bindings: added.
Deprecated Metadata.additional map now marked obsolete/deprecated in C#, Java, and Go.

Changed

PaddleOCR recognition models upgraded to PP-OCRv5 for all 11 script families (arabic, devanagari, tamil, telugu were on v3), improving accuracy.
PDFium upgraded to chromium/7678 (from 7578); C API remains backward-compatible with existing bindings.

4.3.4 - 2026-02-16

Fixed

Node.js keyword extraction fields missing: extractedKeywords, qualityScore, and processingWarnings were silently dropped from results; also renamed the mismatched keywords property to extractedKeywords.
CLI installer resolving benchmark tags as latest release: install.sh could install a benchmark run instead of a real version; it now filters for v-prefixed release tags.

4.3.3 - 2026-02-14

Added

Jupyter notebook image extraction: embedded base64 images (PNG, JPEG, GIF, WebP) are decoded into ExtractedImage structs and OCR’d when configured; SVGs are kept as text.
Markdown data-URI image extraction: data:image/...;base64,... images are decoded and OCR’d when configured; HTTP(S) image URLs are preserved as text markers (no network access).
PaddleOCR 80+ language support via 11 script families (english, chinese, latin, korean, cyrillic, thai, greek, arabic, devanagari, tamil, telugu), downloaded on demand with SHA256 verification, with an engine pool for concurrent multi-language OCR.
Backend-agnostic --ocr-language CLI flag: works across tesseract, paddle-ocr, and easyocr.
Full DOCX extraction pipeline: hierarchical DocumentStructure, per-page splitting, OCR on embedded images, typed metadata fields, style-based heading detection, headers/footers and footnote references, markdown formatting (bold/italic/underline/strikethrough/links), and merged-cell table metadata.

Fixed

LaTeX zero-arg command handling: 35 zero-argument commands (\par, \noindent, etc.) no longer swallow the following {...} group, preventing silent text loss.
Structured-data field mislabeling: field-name matching is now exact, so fields like “width” or “valid” are no longer treated as ID text fields.

4.3.2 - 2026-02-13

Changed

PHP now requires 8.4+: the PHP binding requires PHP 8.4+ (needed for PHPUnit 13.0).

Fixed

Elixir v4.3.1 publish failure: fixed a macOS ARM64 build timeout that had prevented the Elixir package from reaching Hex.pm.

4.3.1 - 2026-02-12

Fixed

Elixir 4.3.0 install failures: fixed a Hex package checksum mismatch (the release shipped with outdated 4.2.10 checksums) across all 8 precompiled NIF binaries.
WASM build failures: fixed missing WebAssembly support in transitive dependencies (ahash, lopdf, rand_core).

4.3.0 - 2026-02-11

Added

Blank page detection: is_blank on PageInfo and PageContent flags near-empty pages (all bindings).
PaddleOCR backend (paddle-ocr feature, ONNX Runtime): PP-OCRv4 models for English, Chinese, Japanese, Korean, German, and French with automatic model download/caching and superior CJK recognition; available in all bindings and via kreuzberg --ocr-backend paddle-ocr.
Structured OCR element output: OcrElement data with bounding geometry, per-element confidence, rotation, and hierarchical levels (word/line/block/page) from both PaddleOCR and Tesseract.
Document structure output: include_document_structure option produces structured DocumentStructure (all bindings).
Native DOC/PPT extraction via OLE/CFB parsing — legacy Office formats no longer require external tools (LibreOffice removed).
musl/Alpine Linux support: prebuilt CLI binaries, Python musllinux wheels, and Node.js native bindings for musl targets, fixing the glibc 2.38+ requirement on older distros (e.g. Ubuntu 22.04).

Fixed

MSG (Outlook) extraction hang on large attachments: fixed indefinite hang by switching to direct CFB parsing; also handles truncated sector tables from some Outlook versions.
Rotated PDF text extraction: PDFs with 90°/270° page rotation previously returned empty content; now extract correctly.
CSV and Excel extraction quality: fixed near-empty/garbled output; both now emit clean delimited text.
XML extraction quality: better handling of namespaced elements, CDATA, and mixed content.
WASM table extraction: tables were silently dropped in Deno/Cloudflare Workers due to a field-name mismatch.
DOCX formatting: extraction now produces formatted markdown (bold/italic/underline/strikethrough/links), correct heading hierarchy, bullet/numbered/nested lists, and inline pipe tables (with cell formatting preserved) instead of plain text.
Typst tables: fixed a parser bug that consumed all remaining content after a #table() call.
PaddleOCR model loading: fixed a ShapeInferenceError on ONNX Runtime 1.23.x and an incorrect detection model filename.
Python: OcrConfig ignoring paddle_ocr_config/element_config, and keyword/additional metadata being dropped from results.
TypeScript/Node.js: PaddleOCR/element config dropped, and ocr_elements missing from results.
Ruby: gem build failures (missing vendored crate), PaddleOCR/element config not parsed, and ocr_elements missing from results.
Go: PdfMetadata deserialization failing when keyword extraction returns object arrays.
C#: keyword extraction data was inaccessible (excluded from serialization).
PHP: document/elements/ocrElements inaccessible, include_document_structure silently ignored, missing OCR-backend management functions, and a page_count key mismatch.
Elixir: config parser dropped several options (include_document_structure, output_format, security_limits, etc.) and lacked document-extractor management functions.

Removed

LibreOffice is no longer required: legacy .doc/.ppt now extract natively, shrinking the full Docker image by ~500-800MB. Users on Kreuzberg <4.3 still need LibreOffice for these formats.
Guten OCR backend references removed; KREUZBERG_DEBUG_GUTEN env var renamed to KREUZBERG_DEBUG_OCR.

4.2.15 - 2026-02-08

Added

Agent Skill for document extraction: Added a SKILL.md following the Agent Skills open standard, with Python/Node.js/Rust/CLI instructions and reference files; works with Claude Code, Codex, Gemini CLI, Cursor, VS Code, and other compatible tools.
Added .docbook (application/docbook+xml) and .jats (application/x-jats+xml) file extension mappings.

Fixed

ODT lists and sections: Documents with bulleted/numbered lists or sections no longer return empty content.
UTF-16 EML: EML files encoded in UTF-16 (LE/BE, with or without BOM) no longer return empty content.
Email attachments: Fixed a metadata serialization bug that overwrote the structured attachments array, causing deserialization failures in Go, C#, and other typed bindings.
WASM office documents: DOCX, PPTX, and ODT now extract correctly in WASM builds.
WASM PDF in non-browser runtimes: PDF extraction now works in Node.js, Bun, and Deno, not just the browser.
Elixir page metadata: Encoding page structure metadata (PageBoundary, PageInfo, PageStructure) to JSON no longer fails.
Pre-built CLI mcp command: Standalone CLI binaries now include the mcp feature, so kreuzberg mcp is available.
PDF error handling: Reverted a v4.2.14 regression; corrupted PDFs again return PdfError::InvalidPdf and password-protected PDFs return PdfError::PasswordRequired instead of silently returning empty results.

Changed

Added security_limits field to all 9 language bindings for parity with the Rust core ExtractionConfig.

4.2.14 - 2026-02-07

Fixed

Excel file-path extraction: .xla (legacy add-in) and .xlsb (binary spreadsheet) files now fall back gracefully when extracted by file path, not just from bytes.

4.2.13 - 2026-02-07

Added

WASM office formats: DOCX, PPTX, RTF, reStructuredText, Org-mode, FictionBook, Typst, BibTeX, and Markdown are now available in the browser/WASM build.
Citation extraction: Added structured extraction for RIS (.ris), PubMed/MEDLINE (.nbib), and EndNote XML (.enw), with authors, DOI, year, keywords, and abstract.
JPEG 2000 OCR: JP2 container and J2K codestream images are now decoded for OCR via a pure-Rust decoder, plus JP2 metadata parsing.
JBIG2 images: Added JBIG2 bi-level image decoding for OCR (common in scanned PDFs), with image/x-jbig2 and .jbig2/.jb2 mappings.
Gzip archives: Added extraction of text from gzip-compressed files (.gz), with decompression size limits to guard against gzip bombs.
JATS and DocBook: Registered the JATS and DocBook extractors in the default registry so these formats now extract.
Added many MIME type and extension mappings for broader format recognition (FictionBook, BibTeX, DocBook, PubMed, Djot, Typst, EPUB, RTF, Org, and more).
Configurable archive security: Added a security_limits field to ExtractionConfig; ZIP, TAR, 7z, and GZIP extractors now enforce configurable limits for archive size, file count, compression ratio, and content size, and ZIP archives are validated for zip bombs before extraction.

Fixed

Typst / Djot recognition: .typ files are now recognized as Typst and .djot files as Djot.
Gzip MIME: application/gzip is no longer rejected by MIME validation.
Case-sensitive MIME validation: Valid MIME types with different casing (e.g. macroEnabled) are no longer rejected.
JPEG 2000 images: .jp2 images are now handled by the image extractor (image/jp2, image/jpx, image/jpm, image/mj2).
YAML: YAML files are no longer rejected; all four YAML MIME variants including the standard application/yaml are accepted.
CLI config: The --config flag now accepts both .yml and .yaml.
.tgz archives: .tgz files are now correctly handled as gzip-compressed TAR rather than raw TAR.
Excel exotic formats: .xlam, .xla, and .xlsb files lacking standard workbook data now return an empty workbook instead of erroring.

4.2.12 - 2026-02-06

Fixed

DOCX lists: Fixed list items missing whitespace between text runs, which merged words together.

4.2.11 - 2026-02-06

Fixed

Python: CLI binary is no longer missing from platform wheels.
OCR fallback: Fixed scanned multi-page PDFs skipping OCR; page count is now passed correctly, and each page is evaluated independently so any scanned page triggers OCR for the document.

4.2.10 - 2026-02-05

Fixed

MIME detection: DOCX/XLSX/PPTX files are no longer misdetected as application/zip when detecting from bytes.
Java: Format-specific metadata (e.g. sheet_count, sheet_names) is no longer missing from getMetadataMap(), and fixed a ClassCastException when deserializing nested generic collections.
Python: Fixed the Windows CLI binary still missing from the wheel.

4.2.9 - 2026-02-03

Fixed

MCP server: Fixed a “Cannot start a runtime from within a runtime” panic when running the MCP server in Docker.
Python: Fixed an “embedded binary not found” error on Windows caused by missing .exe extension handling.
OCR fallback: Fixed scanned PDFs incorrectly skipping OCR; mixed-content PDFs with some scanned pages are now properly OCR’d.

4.2.8 - 2026-02-02

Fixed

Python: Fixed ChunkingConfig serializing wrong field names (max_characters/overlap instead of max_chars/max_overlap).
Java: Fixed an ARM64 SIGBUS crash when reading error details.
Ruby: Fixed a LoadError during native extension compilation caused by a missing rb_sys runtime dependency.

4.2.7 - 2026-02-01

Added

API: Added an OpenAPI schema for the /extract endpoint with full type documentation.
Chunking: Added a unified ChunkingConfig with canonical field names and serde aliases for backwards compatibility.
OCR: Added KREUZBERG_OCR_LANGUAGE="all" to auto-detect and use all installed Tesseract languages.

Fixed

C#: Fixed Attributes deserialization on ARM64 to handle both array-of-arrays and object JSON formats.
Overhauled type definitions across the Elixir, TypeScript, PHP, Ruby, Python, Java, C#, and Go bindings to match the Rust source, correcting many struct/field names, types, and optionality.

Performance

Reduced heap allocations across string-literal fields and the RST, FictionBook, and email extractors, and enabled zero-copy cloning of extracted image data.

4.2.6 - 2026-01-31

Fixed

Python: Restored missing output_format, result_format, elements, and djot_content fields on ExtractionResult.
Python: Chunks now return proper Chunk objects with attribute access instead of dicts.

4.2.5 - 2026-01-30

Fixed

Python: Restored missing OutputFormat/ResultFormat exports that caused ImportError, and fixed Python 3.10 StrEnum compatibility.
PHP: Aligned ImageExtractionConfig, PdfConfig, ImagePreprocessingConfig, and ExtractionConfig with the Rust core and removed phantom parameters.
TypeScript/Node: Restored missing elements field with Element, ElementMetadata, and BoundingBox types.
C#: Fixed enum serialization on .NET 9+.

Added

Node: Added Bun runtime support.

Changed

Achieved PageContent field parity across all language bindings.

4.2.4 - 2026-01-29

Fixed

TypeScript/Node: Restored missing elements field with Element, ElementType, BoundingBox, and ElementMetadata types.
Rust Core: Fixed KeywordConfig failing to deserialize partial configs.
C#: Fixed Element deserialization for the element_based result format.

4.2.3 - 2026-01-28

Fixed

API: Endpoints (/embed, /chunk, others) now reject JSON arrays in request bodies with a 400 instead of malfunctioning.
CLI: --format json now serializes the complete ExtractionResult including chunks, embeddings, images, pages, and elements.
MCP: Tool responses now return the full JSON-serialized ExtractionResult, matching API and CLI output; fixed a config-merge bug that corrupted boolean settings.
Elixir: Added ExtractionConfig.new/0 and new/1; renamed Chunk.text to content for parity.
C#: File-not-found now throws KreuzbergIOException instead of KreuzbergValidationException.
WASM: Fixed initWasm() failing on Cloudflare Workers and Vercel Edge; added an initWasm({ wasmModule }) option for explicit module injection.

4.2.2 - 2026-01-28

Changed

PHP/Go/TypeScript: Removed non-canonical ExtractionConfig and result fields so all bindings match the Rust source.
Ruby/Java: Fixed enable_quality_processing default from false to true to match Rust.

Fixed

Elixir: Fixed force_build: true breaking production installs; source builds now only run in development.
Docker: Fixed “OCR backend ‘tesseract’ not registered” and “Failed to initialize embedding model” errors via tessdata discovery and a persistent model cache.
API: JSON errors now return a proper ErrorResponse; added chunking/embed validation (including overlap < max_characters); EmbeddingConfig.model defaults to the “balanced” preset.
Rust Core: Fixed XLSX out-of-memory on Excel Solver files declaring extreme cell dimensions.

4.2.1 - 2026-01-27

Fixed

Rust Core: Fixed reversed PPTX image page numbers; accept all output-format aliases (plain, text, markdown, md, djot, html); return an IO (not validation) error for file-not-found; log previously silent plugin failures.
Ruby: extract/detect now accept both positional and keyword arguments; renamed image_extraction to images with a backward-compatible alias.
PHP: Renamed fields to canonical names (images, pages, pdfOptions, postprocessor, tokenReduction) and added missing postprocessor/tokenReduction.
Java: Added getImages()/images() aliases for image extraction.
WASM: Added outputFormat, resultFormat, and htmlOptions to the config interface.
Go/Elixir: Added text and md output-format aliases.

Documentation

Added a Kubernetes deployment guide with health checks and troubleshooting.

4.2.0 - 2026-01-26

Added

MCP: Full config parameter support on all tools, enabling complete configuration pass-through.
CLI: Added --output-format (replaces --content-format), --result-format, --config-json, and --config-json-base64 flags.
All bindings: Added outputFormat/output_format (Plain, Markdown, Djot, HTML) and resultFormat/result_format (Unified, ElementBased) fields; Go adds functional options, Java adds Builder methods.
PHP: Added 6 missing config fields (useCache, enableQualityProcessing, forceOcr, maxConcurrentExtractions, resultFormat, outputFormat).

Changed

Configuration precedence is now CLI flag > inline JSON > config file > defaults.

Fixed

Ruby: Fixed batch chunking operations.

BREAKING CHANGES

MCP only (AI-facing, no end-user impact): Removed top-level enable_ocr/force_ocr; use config.ocr.enable_ocr and config.force_ocr. Old names accepted in v4.2 with deprecation warnings.

Deprecated

--content-format CLI flag and KREUZBERG_CONTENT_FORMAT env var deprecated in favor of --output-format / KREUZBERG_OUTPUT_FORMAT (backward compatible).

4.1.2 - 2026-01-25

Added

Ruby: Added Ruby 4.0 support.

Fixed

Ruby: Fixed gem native extension build failure.
Go: Fixed Windows hang caused by an FFI mutex deadlock.

4.1.1 - 2026-01-23

Fixed

PPTX: Fixed extraction failing on shapes without text (e.g. image placeholders).
Added PPSX (PowerPoint Show) and PPTM (macro-enabled) file support.

4.1.0 - 2026-01-21

Added

API: Added POST /chunk endpoint with configurable max_characters, overlap, and trim.
Core: Added Djot markup support (.djot) with parser, structured representation, and YAML frontmatter; Djot output for HTML/OCR; a ContentFormat enum (Plain, Markdown, Djot, HTML) with --content-format; and an element-based result format (Unstructured.io-compatible semantic elements).
All bindings: Added content/result format configuration plus Element, ElementType, ElementMetadata, BoundingBox, and DjotContent types.

Fixed

Python: Restored missing type exports (Element, ElementMetadata, ElementType, BoundingBox, HtmlImageMetadata).
Elixir: Fixed crash when extracting DOCX files with keyword metadata.

4.0.8 - 2026-01-17

Changed

Docker: Migrated images to GitHub Container Registry (ghcr.io/kreuzberg-dev/kreuzberg).

Fixed

C FFI: Fixed empty HtmlConversionOptions serializing as null, causing FFI errors.
Python: Fixed missing _internal_bindings.pyi type stub in wheels.

4.0.6 - 2026-01-14

Fixed

PHP: Fixed runtime panic from unregistered ChunkMetadata and Keyword classes.

4.0.5 - 2026-01-14

Added

Go: Added an automated installer that downloads the correct platform-specific FFI library from releases.

4.0.4 - 2026-01-13

Fixed

Docker: Fixed MissingDependencyError when extracting legacy MS Office formats by adding LibreOffice symlinks and runtime dependencies.

4.0.3 - 2026-01-12

Added

HTML config: Full html_options configuration is now available from config files and all language bindings.

Fixed

Go: Fixed header include path so go get no longer fails with missing-header errors.
C#: Fixed JsonException during keyword extraction; keywords now deserialize as ExtractedKeyword objects.
Distribution: Made the Homebrew tap public to enable brew install kreuzberg-dev/tap/kreuzberg.

4.0.2 - 2026-01-12

Fixed

Go: Fixed module tag format so go get works correctly.
Elixir: Fixed macOS native library extension (.dylib instead of .so).

4.0.1 - 2026-01-11

Fixed

Elixir: Fixed NIF binaries not being published to releases, which broke rustler_precompiled.
Python: Fixed kreuzberg-tesseract missing from source distributions, breaking builds from source.
WASM: Removed a call to the non-existent detectMimeType() API.
Ruby: Updated RBS type definitions to match keyword-argument signatures.

4.0.0 - 2026-01-10

Highlights

First stable release of Kreuzberg v4 — a complete rewrite with a Rust core and polyglot bindings for Python, TypeScript, Ruby, PHP, Java, Go, C#, Elixir, and WebAssembly.

Added

Python FFI error introspection via get_last_error_code() and get_last_panic_context().
PHP custom extractor support, with metadata and tables flowing through to results.
Dynamic Tesseract language discovery from the local installation.

Removed

Removed the v3 legacy Python package and infrastructure.

3.22.0 - 2025-11-27

Fixed

Fixed EasyOCR import error handling.

3.21.0 - 2025-11-05

Added

Rewritten on a Rust core with bindings for Python, TypeScript, Ruby, Java, Go, and C#, plus a REST API and MCP server.

Changed

Architecture restructured around a Rust core with thin language-specific wrappers.

Security

Sandboxed subprocess execution, input validation, and memory safety via Rust.

Performance

Streaming PDF extraction, SIMD optimizations, and ONNX Runtime for embeddings.

3.20.2 - 2025-10-11

Fixed

Fixed missing optional dependency errors in GMFT extractor.

3.20.1 - 2025-10-11

Maintenance release (internal changes only).

3.20.0 - 2025-10-11

Added

Python 3.14 support.

Changed

Migrated HTML extractor to html-to-markdown v2.

3.19.1 - 2025-09-30

Fixed

Fixed Windows Tesseract 5.5.0 HOCR output compatibility.

3.19.0 - 2025-09-29

Fixed

Fixed Tesseract PSM handling and aligned sync/async OCR pipelines.
Fixed ValidationError handling in batch processing.

3.18.0 - 2025-09-27

Added

API server configuration via environment variables.
Auto-download of missing spaCy models for entity extraction.

Fixed

Fixed HOCR parsing issues.

3.17.0 - 2025-09-17

Added

Token reduction for text optimization with streaming support.

Fixed

Fixed excessive markdown escaping in OCR output. (#133)

3.16.0 - 2025-09-16

Added

Enhanced JSON extraction with schema analysis and custom field detection.

Fixed

Fixed HTMLToMarkdownConfig not being exported in the public API.
Fixed Windows-specific path issues.

3.15.0 - 2025-09-14

Added

Comprehensive image extraction support.
Polars DataFrame and PIL Image serialization for API responses.

Fixed

Fixed TypeError with unhashable dict in API config merging.

3.14.0 - 2025-09-13

Added

DPI configuration system for OCR processing.

Changed

Enhanced API with 1GB upload limit and comprehensive OpenAPI documentation.
Completed pandas to polars migration.

3.13.0 - 2025-09-04

Added

Runtime configuration API via query parameters and headers.
OCR caching for EasyOCR and PaddleOCR backends.

Fixed

Fixed Tesseract TSV output and table extraction.
Fixed UTF-8 encoding handling across document processing.
Fixed regression in PDF extraction and XLS file handling.

3.12.0 - 2025-08-30

Added

Multilingual OCR support in Docker images with flexible backend selection.

Fixed

Fixed naming conflict in CLI config command.

3.11.1 - 2025-08-13

Fixed

Faster startup by deferring numpy import until image processing.

3.11.0 - 2025-08-01

Fixed

Fixed image extractor async delegation.
Fixed timezone handling in spreadsheet metadata.

3.10.0 - 2025-07-29

Added

PDF password support through new crypto extra feature.

3.9.0 - 2025-07-17

Maintenance release (internal changes only).

[3.8.0] - 2025-07-16

Maintenance release (internal changes only).

3.7.0 - 2025-07-11

Added

MCP server enabling Claude integration with document extraction.

Fixed

Fixed chunk parameters to prevent overlap validation errors.

3.6.0 - 2025-07-04

Added

Language detection integrated into the extraction pipeline.

Changed

Entity extraction now uses spaCy, replacing gliner.

3.5.0 - 2025-07-04

Added

Language detection with configurable backends.
Full synchronous support for PaddleOCR and EasyOCR backends.

Changed

Python 3.10+ now required (3.9 support dropped).

3.4.0 - 2025-07-03

Added

REST API (Litestar) for web-based document extraction.

Fixed

Fixed race conditions in GMFT and Tesseract caching.

3.3.0 - 2025-06-23

Added

GMFT table extraction via an isolated process wrapper.
CLI support.
Pure synchronous extractors without anyio dependencies.
Document-level caching with per-file locks and parallel batch processing.

Fixed

Fixed Windows-specific multiprocessing failures.
Fixed file existence validation in extraction functions.

Changed

Faster cache serialization (msgpack replaces msgspec JSON, ~5x).

3.2.0 - 2025-06-23

Added

GPU acceleration for OCR and ML operations.
Multiple language support for EasyOCR.

Fixed

Fixed EasyOCR byte string issues.

3.1.0 - 2025-03-28

Added

GMFT (Give Me Formatted Tables) support for vision-based table extraction.

Changed

Image extraction now non-optional in results.

3.0.0 - 2025-03-23

Added

Chunking for document segmentation.
Hooks system for pre/post-processing.
OCR backend abstraction with EasyOCR and PaddleOCR support and multi-language OCR.

Fixed

Fixed Windows error message handling.
Fixed PaddleOCR integration issues.