Releasesby Glyphunstructured 0.22.16
0.22.16 Enhancements - Formula markdown export (element_to_md / elements_to_md): New keyword-only formula_markdown_style ("auto", "display_math", "plain"; default "auto"). In "auto", display math ($ ... $) is used only when the text looks like notation (heuristic score) and contains no $/$ (avoids breaking Markdown and noisy OCR captions).
Also shipped Apr 3, 2026Unstructured in April 2026 →
- [Preview] Embedded Dives: MotherDuck Dives can now be embedded directly in your customer-facing applications
- [Preview] Embedded Dives: MotherDuck Dives can now be embedded directly in your customer-facing applications. Create a Dive, generate an embed session, and drop an iframe into your app. Your end-users can explore data in real time, powered by MotherDuck. Embedded Dives currently require a Business plan.
AI shortcuts in the Neon Docs, Postgres extension updates, Azure region deprecation, Mastra and Vercel AI SDK guides, and more release - Apr 03, 2026
AI assisted shortcuts in the Neon Docs. Copy page in the header opens a menu on every Neon Docs page copy the page as Markdown, or open it in ChatGPT or Claude ...
What's new in Sigma
AI New model used for Snowflake AI provider If you use a warehouse-hosted AI provider with a Snowflake connection, and your Snowflake account is hosted on AWS, Sigma now uses claude-sonnet-4-5 as the LLM. For more details, see Configure a warehouse-hosted model as AI provider .
Guides and Surveys showing up in cohort page's _Used by_ tab
What it is: The "Used by" tab on the Cohorts page shows which assets reference a given cohort, including charts, other cohorts, experiments, and flags. Guides and Surveys now appear in that list as well. Previously, those asset types weren't included, leaving the dependency picture incomplete.
NVIDIA LLM provider now available
NVIDIA is now a supported LLM provider for the Voice Agent API. The following model is available in the Standard pricing tier: nemotron-3-nano-30B-A3B , Nemotron 3 Nano 30B A3B provides cost efficiency with high accuracy for targeted agentic tasks.
LlamaIndex 0.14.20
Release Notes [2026-04-03] llama-index-agent-agentmesh [0.2.0] fix vulnerability with nltk ( #21275 ) llama-index-callbacks-agentops [0.5.0] chore(deps): bump the uv group across 50 directories with 2 updates ( #21164 ) chore(deps): bump the uv group across 24 directories with 1 update ( #21219 ) chore(deps): bump the uv group across 21 directories with 2...
More from unstructuredEvery unstructured release →
unstructured 0.22.12
- mem: exclude unused spaCy pipeline components to reduce model memory - fix: pdfminer drops extractable text
unstructured 0.22.10
- fix(chunking): preserve nested table structure in reconstruction - Replace lazyproperty with functools.cached_property - mem: reduce PaddleOCR rec_batch_num from 6 to 1 - fix: isolate Table elements in pre-chunks - feat(chunking): repeat table headers on continuation chunks
unstructured 0.22.18
- fix(chunking): preserve semantic headers in carried table chunks - feat: add page number support to v1 html partition
unstructured 0.22.6
- fix(deps): Update security updates [SECURITY] - fix: Self-contained script for version extraction in release CI
unstructured 0.22.4
- feat: add create_file_from_elements() to re-create document files from elements - Bump dependencies - fix: avoid O(N²) re-scanning in _patch_current_chars_with_render_mode - add check if libmagic fails - Adds Form Element - feat: audio speech to text partition - Add a check for complex pdfs - chore: disable fail-build on Anchore container scan - feat...
unstructured 0.22.21
- feat: add option to skip table chunking
unstructured 0.22.20
- Fix fixtures update CI to regenerate markdown - fix(deps): upgrade vulnerable transitive dependencies [security] - feat: add GHA workflow to build opencv wheels without ffmpeg - Enable vertical text detection for rotated images
unstructured 0.22.22
Security - Replace PyPI opencv wheels with ffmpeg-free builds in Docker image: After uv sync, the Dockerfile now substitutes all PyPI opencv-python variants with a source-built opencv-contrib-python-headless wheel compiled with WITH_FFMPEG=OFF, eliminating 14 bundled ffmpeg CVEs.