NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #4982 most downloaded on PyPI
A library for performing inference using trained models.
Latest stable
1.6.17
Last release 25 days ago
28 Oct 2026
Release timing varies
gaps range from 8 days to 4 months
Nearly every release is documented
notes for 60 of the last 60 stable releases
2 versions withdrawn
withdrawn after publishing
4 years old
127 releases · first in 2022
Drag across the ruler to compare, or tap a release to open it.
perf(layout): Process multi-frame images one frame at a time by @CyMule in https://github.com/Unstructured-IO/unstructured-inference/pull/520
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.16...v1.6.17
fix(deps): bump certifi to 2026.10.19 by @lawrence-u10d in https://github.com/Unstructured-IO/unstructured-inference/pull/529
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.14...v1.6.16
Your coding agent can read these notes before it upgrades. Set up the MCP server →
perf(layout): Use linear memory when cleaning up large region groups by @CyMule in https://github.com/Unstructured-IO/unstructured-inference/pull/522
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.13...v1.6.14
fix: render filled PDF form field values (init_forms) by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/515
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.12...v1.6.13
convert_pdf_to_image now calls init_forms() so AcroForm/XFA field values (text typed into fillable fields) are painted into the rendered page image instead of being silently dropped.fix: render rotated PDF pages consistently with extracted text by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/514
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.11...v1.6.12
convert_pdf_to_image no longer re-applies the page /Rotate on top of pypdfium2's render (which already produces the display frame), so the rendered image stays in the same coordinate frame as pdfminer's extracted text. For pages with a non-zero /Rotate, the dominant text orientation is detected and, when a single 90-degree orientation clearly dominates (configurable via PDF_ROTATION_DOMINANT_ANGLE_THRESHOLD), the image is rotated so its text is upright. The applied correction is exposed as pdf_rotation_correction in the page image metadata so downstream consumers can keep extracted coordinates aligned.feat: Track the table extraction method by @vladimir-kivi-ds in https://github.com/Unstructured-IO/unstructured-inference/pull/513
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.10...v1.6.11
table_extraction_method field to LayoutElements and LayoutElement to track which algorithm produced a table (grid, tatr, vlm).Allow Python 3.13 support by @CyMule in https://github.com/Unstructured-IO/unstructured-inference/pull/511
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.9...v1.6.10
Update Python 3.11 support by @CyMule in https://github.com/Unstructured-IO/unstructured-inference/pull/510
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.8...v1.6.9
Reject oversized PDF renders before bitmap allocation by @CyMule in https://github.com/Unstructured-IO/unstructured-inference/pull/509
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.7...v1.6.8
fix: fix rare occasion of initialization error by converting LazyDict into dict by @badGarnet in https://github.com/Unstructured-IO/unstructured-infer
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.6...v1.6.7
get_model now materializes LazyDict model configs into a plain dict before
unpacking into initialize(**...). Uses __iter__ + __getitem__ to avoid
depending on Mapping.keys(), which has been observed to fail at **
unpacking with "argument after ** must be a mapping, not LazyDict" in some
deployment environments.Relax pandas and numpy constraints by @vladimir-kivi-ds in https://github.com/Unstructured-IO/unstructured-inference/pull/507
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.5...v1.6.6
Unit test for pdf page rotation by @vladimir-kivi-ds in https://github.com/Unstructured-IO/unstructured-inference/pull/505
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.4...v1.6.5
pdf_rotation in page.image_metadata so downstream consumers can check page rotation after the page image is freedconvert_pdf_to_imagefix(deps): upgrade vulnerable transitive dependencies [security] by @lawrence-u10d in https://github.com/Unstructured-IO/unstructured-inference/pull/5…
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.2...v1.6.4
/Rotate metadata during page rendering - pypdfium2's page.render() ignores the flag, producing sideways images for rotated pagesmem: free intermediate arrays during YoloX inference by @KRRT7 in https://github.com/Unstructured-IO/unstructured-inference/pull/496
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.6.0...v1.6.2
dpi an explicit parameter on convert_pdf_to_image (default 200) instead of reading from config internally, enabling unstructured to use this as the single source of truth for PDF renderingmem: use np.full() instead of np.ones() * scalar in YoloX preprocess by @KRRT7 in https://github.com/Unstructured-IO/unstructured-inference/pull/497
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.5.3...v1.6.0
huggingface-hub lower bound from >=1.4.1 to >=0.22.0 (the >=1.4.1 was an artifact of the uv migration and broke compatibility with transformers<5.0)chore: add Renovate config (python-uv preset) by @lawrence-u10d in https://github.com/Unstructured-IO/unstructured-inference/pull/472
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.5.2...v1.5.3
Enable trusted publishing by @PastelStorm in https://github.com/Unstructured-IO/unstructured-inference/pull/471
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/v1.5.1...v1.5.2
feat: bump pypdfium minimum and use context manager by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/460
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/1.1.8...1.2.0
Update OpenCV-Python to 4.13.0.90 to squash ffmpeg vulnerability CVE-2023-6605
Use inference_config to set default rendering DPI
Constrain urllib3 to urllib3>=2.6.0 to address CVE-2025-66471 and CVE-2025-66418
chore(deps): Bump several depedencies to resolve open high CVEs
Add NotImplementedError when trying to single index a TextRegions, reflecting the fact that it won't behave correctly at the moment.
TextSource to track where the text of an element came from__post_init__ of TextRegions and LayoutElement slightly to automate initializationzoom_image (codeflash)cells_to_html for an 8% speedup in some cases (codeflash)outputs_to_objects for an 88% speedup in some cases (codeflash)DetrImageProcessor with env variablesglobal to store shared variablesfeat: add thread lock to prevent racing condition when instantiating singletons
DetrImageProcessor with env variablesBREAKING stop supporting python 3.8
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.10...1.0.2
unpin numpy and raise min for onnx higher to be compatible by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/416
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.9...0.8.10
numpy and bump minimum for onnxruntime to be compatible with numpy>=2feat: make PageLayout.elements a cached property by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/414
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.8...0.8.9
Nothing published for this version
feat: add password param for protected pdfs by @pprados in https://github.com/Unstructured-IO/unstructured-inference/pull/392
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.6...0.8.7
feat: add back source attribute for backward compatibility by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/407
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.5...0.8.6
source to TextRegions and LayoutElements for backward compatibilityremove pdfplumber but include pdfminer-six==20240706 to update pdfminer
remove pdfplumber but include pdfminer-six==20240706 to update pdfminer
pdfplumber but include pdfminer-six==20240706 to update pdfminerFeat/add more attributes to layoutelements by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/404
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.3...0.8.4
text_as_html and table_as_cells to LayoutElements class as new attributessource attribute from TextRegions and LayoutElements with an array attribute sourcesfix: drop layoutparser model support
layoutelement.from_lp_textblock() and related tests as it's not supportedlayoutparser liblayoutelement.from_lp_textblock() and related tests as it's not usedlayoutparser libREADME.md to remove layoutparser model zoo support notefix: fix bugs in data structure by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/402
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.8.1...0.8.2
TextRegions.from_list triggers IndexErrorLayoutElements the class id mapping is no properly
updatedfix: fix list index out of range error caused by calling LayoutElements.from_list() with empty list
Remove chipper model by @vangheem in https://github.com/Unstructured-IO/unstructured-inference/pull/395
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.41...0.8.0
fix: fix incorrect type casting with higher versions of numpy when substracting a float from an int array
numpy when substracting a float from an int arrayNone when calling LayoutElements.as_list()Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.39...0.7.40
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.39...0.7.40
update version for release by @badGarnet in https://github.com/Unstructured-IO/unstructured-inference/pull/389
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.38...0.7.39
LayoutElements classNothing published for this version
build(deps): update setup-python gha by @qued in https://github.com/Unstructured-IO/unstructured-inference/pull/360
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.36...0.7.37
numpy array to store layout elements' information in one single LayoutElements
object instead of using a list of LayoutElementfix: add input parameter validation to fill_cells() when converting cells to html
fill_cells() when converting cells to htmlFix syntax for generated HTML tables
* Reduce excessive logging
BREAKING CHANGE: removes legacy detectron2 model
refactor: remove all code related to filling inferred elements text from embedded text (pdfminer).
refactor: remove all cid related code that was originally added to filter out invalid pdfminer text
cid related code that was originally added to filter out invalid pdfminer textfix: table transformer predictions are now removed if confidence is below threshold
feat: allow table transformer agent to return table prediction in not parsed format
chore: update pin on onnxruntime dependency.
onnxruntime dependency.feat: add a set of new ElementTypes to extend future element types recognition
ElementTypes to extend future element types recognitionunstructured_inference.models.base.register_new_model functionfix: replace Rectangle.is_in() with Rectangle.is_almost_subregion_of() when filling in an inferred element with embedded text
Rectangle.is_in() with Rectangle.is_almost_subregion_of() when filling in an inferred element with embedded textinstall-detectron2 from the MakefileUNSTRUCTURED_DEFAULT_MODEL_INITIALIZE_PARAMS_JSON_PATH to int typeFix: text_as_html attribute appearing in every element metadata by @mpolomdeepsense in https://github.com/Unstructured-IO/unstructured-inference/pull/
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.23...0.7.24
text_as_html element attribute only if text attribute contains HTML tags.fix: added handling in UnstructuredTableTransformerModel for if recognize returns an empty list in run_prediction.
UnstructuredTableTransformerModel for if recognize returns an empty
list in run_prediction.Feat/chipper repetitions by @ajjimeno in https://github.com/Unstructured-IO/unstructured-inference/pull/295
Full Changelog: https://github.com/Unstructured-IO/unstructured-inference/compare/0.7.20...0.7.21
PageLayout object, lack image_metadata and other attributes that are required for downstream processing; this fix also reduces the memory overhead of using chipper modelchipper-v3: improved table prediction
refactor: remove all OCR related code
refactor: remove all image extraction related code
refactor: remove all pdfminer related code
pdfminer related codebug: Allow supplied ONNX models to use label_map dictionary from json file
One column per quarter.