## What's Changed * fix: first table chunk preserve col/row span by @badGarnet in https://github.com/Unstructured-IO/unstructured/pull/4343 **Full Changelog**: https://github.com/Unstructured-IO/unstructured/compare/0.22.22...0.22.23

OpenCV Python contrib-headless wheels built from source with WITH_FFMPEG=OFF. These wheels eliminate bundled ffmpeg CVEs present in the stock PyPI wheels. Built against cgr.dev/chainguard/wolfi-base:latest with Python 3.12. The contrib-headless variant provides the full cv2 API (core + contrib modules, no GUI), so a single wheel can satisfy opencv-python, opencv-python-headless, opencv-contrib-python, and opencv-contrib-python-headless in downstream Dockerfiles. **Source version:** opencv-contrib-python-headless==4.12.0.88 **Build flags:** CMAKE_ARGS='-DWITH_FFMPEG=OFF' ENABLE_CONTRIB=1 ENABLE_HEADLESS=1

## What's Changed * mem: exclude unused spaCy pipeline components to reduce model memory by @KRRT7 in https://github.com/Unstructured-IO/unstructured/pull/4296 * fix: pdfminer drops extractable text by @qued in https://github.com/Unstructured-IO/unstructured/pull/4310 **Full Changelog**: https://github.com/Unstructured-IO/unstructured/compare/0.22.10...0.22.12

## What's Changed * feat: custom fallback for language detection by @claytonlin1110 in https://github.com/Unstructured-IO/unstructured/pull/4238 * Add Github action for time regressions by @aadland6 in https://github.com/Unstructured-IO/unstructured/pull/4261 * fix: relax lower bound for pdfminer.six by @badGarnet in https://github.com/Unstructured-IO/unstructured/pull/4262 ## New Contributors * @claytonlin1110 made their first contribution in https://github.com/Unstructured-IO/unstructured/pull/4238 * @aadland6 made their first contribution in https://github.com/Unstructured-IO/unstructured/pull/4261 **Full Changelog**: https://github.com/Unstructured-IO/unstructured/compare/0.21.2...0.21.5

## 0.21.0 ### Fixes - **Replace NLTK with spaCy to remediate CVE-2025-14009**: NLTK's downloader uses `zipfile.extractall()` without path validation, enabling RCE via malicious packages (CVSS 10.0, no patch available). spaCy models install as pip packages, eliminating the vulnerable downloader entirely.

Release 0.20.2

## What's Changed * feat: add group_elements_by_parent_id utility function by @MkDev11 in https://github.com/Unstructured-IO/unstructured/pull/4207 * Preserve newlines in Table and TableChunk elements during PDF partitioning by @eureka928 in https://github.com/Unstructured-IO/unstructured/pull/4214 * Fix: make pdf image dpi consistent by @badGarnet in https://github.com/Unstructured-IO/unstructured/pull/4217 * chore: bump dependencies for 0.18.34 by @luke-kucing in https://github.com/Unstructured-IO/unstructured/pull/4221 * feat: increase PIL's max image pixel value for pdf partition by @badGarnet in https://github.com/Unstructured-IO/unstructured/pull/4220 * Migrate to uv by @PastelStorm in https://github.com/Unstructured-IO/unstructured/pull/4226 * Fix ARM64 paddlepaddle image builder bug by @PastelStorm in https://github.com/Unstructured-IO/unstructured/pull/4228 * Fix Docker ARM64 image failure, use 8-core github runners by @PastelStorm in https://github.com/Unstructured-IO/unstructured/pull/4232 * Fix ARM64 image issues by @PastelStorm in https://github.com/Unstructured-IO/unstructured/pull/4233 **Full Changelog**: https://github.com/Unstructured-IO/unstructured/compare/0.18.32...0.19.3

### Enhancement - Optimize `clean_extra_whitespace_with_index_run` (codeflash) - Optimize `recursive_xy_cut_swapped` (codeflash) - Optimize `_DocxPartitioner._parse_category_depth_by_style_name` (codeflash) - Optimize `VertexAIEmbeddingEncoder._add_embeddings_to_elements` (codeflash) - Optimize `ngrams` (codeflash) - Optimize `stage_for_datasaur` (codeflash)