Changelog for Python

General

• Added
  Adds in-process local model support for structured extraction without an external model server. (L#NAVI-86)
• Added
  Adds a no-sampling decision-readout primitive (LLAMAW_Decide/LLAMAW_TokenizeSingle) to the in-process llama.cpp wrapper. (#3645)
• Added
  Adds an opt-in (enableComplexFieldGrounding) path that scores object/array-row fields through the document decision model’s real JSON context instead of a synthesized heuristic sentence; accuracy impact vs. the existing heuristic isn’t yet independently measured. (#3647)
• Added
  Adds opt-in local decision-model backends for document classification and splitting. (#3644, #3648, #3649)
• Added
  Adds continuation links for paragraphs and tables split across pages in the JSON export and merged Markdown output. (J#NAVI-74)
• Added
  Adds dispose() to editors, pages, metadata, and collections, fixing native-resource leaks. (#3422)
• Added
  Adds highlighted-word metadata and preserves highlights in Markdown and JSON exports. (#3315)
• Added
  Adds llama.cpp model discovery for configurable local vision models. (L#NAVI-73)
• Added
  Adds optional partial redaction that preserves the final logical group of detected sensitive values. (#3597)
• Added
  Adds support for the gpt-6-luna OpenAI vision model. (#3487)
• Added
  Adds Vision.Decide, a generic decision-model API that lets callers submit their own state and questions JSON directly to the native decision readout. (#3646)
• Added
  Adds text from small pictures and extracted chart data to Vision JSON output. (#3521)
• Changed
  Improves document processing performance with SIMD-accelerated image preprocessing and model decoding. (#3327)
• Changed
  Improves AI extraction reliability by retrying an OpenAI request once when it times out, instead of failing. (#3580)
• Changed
  Improves chart value accuracy by keeping axis titles and legends inside the region sent to the vision model. (#3322(opens in a new tab))
• Changed
  Improves document parsing: Detects code blocks, recognizes numbered section headings, and removes duplicated OCR text. (#3686)
• Changed
  Improves Markdown export: plain captions, no generated chart summaries, and valid image syntax for pictures. (#3637)
• Changed
  Improves SVG rendering security. (L#NAT-691)
• Changed
  Improves table recognition speed when tables are processed on the CPU. (#3582)
• Changed
  Improves table extraction by preserving cell text on rotated pages, recovering missing rows in long tables, and detecting tables missed in the initial page analysis. (#3488)
• Changed
  Changes the default OpenAI vision model to gpt-6-luna. (#3707)
• Changed
  Changes the default per-attempt OpenAI network timeout from 600 to 300 seconds, so a stalled AI call is retried sooner. (#3777)
• Removed
  Removes the embedded OpenSSL; local VLM model downloads now use the operating system TLS stack and proxy settings. (#3530)
• Fixed
  Fixes an issue where prompts were truncated in the VLM generation path. (#3645)
• Fixed
  Fixes a macOS crash at process exit after running the local in-process vision language model. (L#NAVI-85)
• Fixed
  Fixes process crashes when several threads register a license key or evaluate a licensed feature at the same time. (L#GDPIC-1655)
• Fixed
  Fixes an issue where trial-mode feature traces were lost and trial initialization ran more than once when licensed features were first used concurrently. (L#GDPIC-1655)
• Fixed
  Fixes an issue where the licensing banner and licensed feature list were printed to the console after a production license key was registered. (L#GDPIC-1622)
• Fixed
  Fixes an issue where AI data extraction failed with a misleading error when a schema field had an empty list of allowed values. (L#DWS-806)
• Fixed
  Fixes an issue where AI extraction failed on very large schemas. Properties are now extracted in groups. (#3580)
• Fixed
  Fixes an issue where AI parsing failed a page when the provider’s content filter stopped the VLM reply. Affected zones retain their OCR text. (#3580)
• Fixed
  Fixes AI reader requests failing with Claude Opus 5 and 5.5 and OpenAI gpt-6-sol, gpt-6-astra, and gpt-6.1-sol. (L#NAVI-169)
• Fixed
  Fixes crashes during concurrent document processing with GPU acceleration on macOS. (#3769)
• Fixed
  Fixes CUDA acceleration on Linux systems with glibc older than 2.38, such as Debian 12 and Ubuntu 22.04. (L#NAVI-165)
• Fixed
  Fixes document processing failing on pages that need a targeted re-OCR when the OCR engine is set to PaddleOCR. (#3579)
• Fixed
  Fixes an issue where editors closed their document twice and leaked native handles when released. (#3422)
• Fixed
  Fixes malformed Markdown when extracted text contains less-than signs before HTML tables. (#3473)
• Fixed
  Fixes page rendering in document processing to apply the page’s rotation, as the DocProcessor render command does. (#3693)
• Fixed
  Fixes PDF 2.0 saves to preserve existing XMP and migrate standard Info fields when XMP is absent. (#3091)
• Fixed
  Fixes signing encrypted PDFs without losing protection and opening certificate-encrypted PDFs. (#3091)
• Fixed
  Fixes macOS installation and native-library compatibility; the Python SDK now supports macOS 14 or newer. (#3813)
• Fixed
  Fixes VLM-based extraction failing or misreading text on high-resolution scans and with newer models. (#3382)

Barcode

• Added
  Adds USPS Intelligent Mail Barcode (IMb) recognition to GdPictureImaging with the BarcodeIntelligentMailReader methods. The best quality mode reads barcodes at any angle. Adds GdPictureStatus.BarcodeWrongState, returned when the native barcode reader fails internally. (L#GDBC-498)
• Changed
  Improves QR code recognition on scans with smudged timing patterns or a squeezed band of rows. (#3542)
• Changed
  Improves recognition of QR codes on scans that squeeze rows hard, damage a corner square, or crop the code tightly. (#3545)

Imaging

• Added
  Adds TWAIN methods to read the extended image information (DAT_EXTIMAGEINFO) a scanner reports for each acquired image, including hardware-detected patch codes. (J#GDPIC-1614)
• Changed
  Optimizes the raster operators used by DrawGdPictureImageOP and DrawFilledRectangleOP with SIMD. (#3327)
• Fixed
  Fixes 32-bit processes reporting a production license as trial and printing the evaluation banner. (J#GDPIC-1663)
• Fixed
  Fixes loading of the Linux native libraries on glibc 2.34 systems such as Amazon Linux 2023. (#3514)
• Fixed
  Fixes a memory leak occurring on every scanned page when TWAIN barcode detection is enabled. (J#GDPIC-1614)
• Fixed
  Fixes form field detection crashing on CPU-only machines when the exported model uses float16 precision. (#3650)
• Fixed
  Fixes loading of standalone JBIG2 (.jb2) image files, including multipage files. (L#GDPIC-1656)
• Fixed
  Fixes an issue where MP4, MOV, and other video files were detected as HEIF images and failed to open with a generic error. (#3091)

OCR

• Added
  Adds automatic handwriting layout recovery that reconstructs under-segmented handwritten regions before extraction. (#3374)
• Added
  Adds automatic per-page routing (VisionEngine.Auto) that picks the cheapest suitable extraction pipeline per page. (#3296)
• Added
  Adds GPU-accelerated Paddle OCR with automatic fallback and hybrid Tesseract routing for vision document processing. (#3317)
• Changed
  Improves automatic OCR language detection. (L#NAVI-115)
• Changed
  Improves automatic OCR language selection. (L#NAVI-123)
• Changed
  Improves OCR accuracy on GPU machines. (L#NAVI-126)
• Changed
  Improves automatic OCR language detection speed. (#3566)
• Changed
  Improves bold detection on scanned pages by comparing each word’s stroke weight with same-size text on the page. (#3641)
• Changed
  Improves document splitting accuracy on merged files that contain rotated scanned pages. (#3497)
• Changed
  Improves OCR language detection on scanned pages, identifying more languages correctly in less time. (#3710)
• Changed
  Improves OCR on low-contrast scans. (L#NAVI-108, L#NAVI-109)
• Changed
  Improves OCR processing speed when using GPU acceleration. (#3299)
• Changed
  Improves OCR speed on GPU by checking uncertain words in parallel. (#3566)
• Changed
  Improves page orientation detection on scanned pages. (#3710)
• Changed
  Improves text extraction from scanned pages that carry a few typed or stamped words by reading the whole page with OCR. (#3710)
• Changed
  Keeps the text of table regions that the recognized cell structure doesn’t account for: Words inside a detected table but outside every recognized cell (total lines under the last row, lines merged into a neighboring row) are now exported as text after the table instead of being dropped. (#3342)
• Changed
  Makes document conversion reproducible: Reading order, spatial lookups, and word ordering no longer depend on the order in which internal collections happen to enumerate, so converting the same document twice produces the same text. (#3343)
• Changed
  Changes document data extraction to ignore a PDF’s invisible OCR text layer and read those pages with its own OCR. (#3628)
• Changed
  Enables fully local structured extraction: Extract with provider local and no endpoint now runs the model in-process, downloading the default local model on first use, instead of requiring an OpenAI-compatible server. (L#NAVI-166)
• Changed
  Optimizes OCR processing speed on scanned pages. (#3710)
• Changed
  Optimizes text detection speed in OCR and language detection, most on GPU. (J#NAVI-218)
• Changed
  Updates the in-process local model runtime to llama.cpp 0.5.0, which adds support for newer local models such as Qwen3.8 (models with multitoken prediction layers). (L#NAVI-167)
• Fixed
  Fixes recognized text being dropped from the output when it falls in a region no recognition stage produced content for. (#3343)
• Fixed
  Fixes an issue where label and value words on the same row of a scanned page could be emitted out of order. (L#NAVI-84, #3343)
• Fixed
  Fixes a crash when OCR with several languages reads a word differently in each language. (#3786)
• Fixed
  Fixes automatic OCR language detection misreading Latin-script pages as Cyrillic when Ocr.maxScripts is above 1. (L#NAVI-148)
• Fixed
  Fixes making PDFs searchable in Danish, Swedish, Norwegian, and most other languages, which were read as English. (L#NAPY-24)
• Fixed
  Fixes multizone OCR returning no words for a zone that shares a text line with an earlier zone. (#3609)
• Fixed
  Fixes OCR failing on every page when the OCR engine is pinned to Paddle. (#3423)
• Fixed
  Fixes thousands separators in amounts read from faint scans. (L#NAVI-135)

Office

• Added
  Adds support for the STYLEREF field when converting Word documents to PDF, so a header or footer prints the text of the relevant paragraph — typically the current heading — on each page instead of repeating the value cached in the document. (J#MAMBO-2116)
• Added
  Adds Hebrew and Arabic letter list numbering to DOCX-to-PDF conversion. (J#MAMBO-2129)
• Changed
  Improves the precision of cell and paragraph shading when converting documents to PDF and SVG. (J#MAMBO-2126)
• Fixed
  Fixes HTML and Markdown export of Office documents and images, which failed instead of converting them to PDF first. (#3091)
• Fixed
  Fixes missing rows and incorrect row heights when converting presentation tables to PDF. (J#MAMBO-2126)
• Fixed
  Fixes Excel workbooks failing to open when a table style omits its optional display settings. (#3528)
• Fixed
  Fixes Excel workbooks with damaged internal references failing to open when they could be repaired. (#3528)
• Fixed
  Fixes MSG files rendering as binary garbage when their compound file directory is fragmented. (J#DV-759)
• Fixed
  Fixes a paragraph after a floating table being pushed below the table by unneeded automatic spacing. (L#MAMBO-2131)
• Fixed
  Fixes a table of contents entry or other field result rendering entirely in the font of its first character. (L#MAMBO-1911)
• Fixed
  Fixes field results and field codes being swapped when converting legacy Word documents to Word format. (L#MAMBO-1911)
• Fixed
  Fixes Markdown-to-PDF conversion producing raster pages with no extractable text. (L#GDPIC-1660)
• Fixed
  Fixes multilevel list numbers being scrambled (6.1 shown as 16.) in right-to-left DOCX-to-PDF conversion. (J#MAMBO-2129)
• Fixed
  Fixes pictures placed in a line of text being pinned to a corner when converting legacy Word documents. (L#MAMBO-1911)
• Fixed
  Fixes PowerPoint tables using the “No Style, No Grid” table style being rendered with black cell borders. (L#MAMBO-2132)
• Fixed
  Fixes regular text rendered in a heavy weight when converting Office documents that use static multiweight font families, such as Lato. (#3699)
• Fixed
  Fixes right-to-left text order and column-break layout when converting Word documents to PDF. (J#MAMBO-2045)
• Fixed
  Fixes right-to-left text running under floating text boxes and images instead of wrapping around them in DOCX to PDF. (L#MAMBO-2130)
• Fixed
  Fixes shapes with a picture or texture fill disappearing when converting PowerPoint presentations to PDF. (J#MAMBO-2121)
• Fixed
  Fixes table columns in Word documents being too narrow for their longest word, so words no longer break mid-word. (L#MAMBO-2131)
• Fixed
  Fixes Word documents failing to open with InvalidParameter when a table cell spans zero grid columns. (L#MAMBO-2136)
• Fixed
  Fixes Word table cells with zero cell margins being rendered with the table’s default margins. (L#MAMBO-2131)
• Fixed
  Fixes words being repeated or truncated in a table of contents rendered with small capitals. (L#MAMBO-1911)

PDF

• Changed
  Improves PDF encryption to use random AES initialization vectors and keys and the system cryptographic provider. (#3665)
• Changed
  Improves heading detection for document titles in Markdown output. (#3567)
• Changed
  Improves table detection in PDF layout analysis. (#3567)
• Changed
  Improves detection of ruled tables with stacked records, merged cells, or shaded rows in PDF layout analysis. (#3520)
• Changed
  Improves font handling for specific document configurations. (J#PDFA-114)
• Changed
  Improves PDF-to-HTML conversion speed and memory use for documents with many text runs. (#3631)
• Changed
  Improves text extraction fidelity for ligature glyphs, underlines, and clipped text on rotated PDF pages. (#3488)
• Changed
  Improves text extraction for fonts that identify characters by glyph name. (#3653)
• Changed
  Improves the accuracy of heuristic document layout detection. (#3377)
• Fixed
  Fixes HTML and PDF export replacing the underlying error with a generic one, hiding the real cause of a failed export. (#3091)
• Fixed
  Fixes annotation text such as contents and author reading back garbled when stored as UTF-16 or non-ASCII. (#3505)
• Fixed
  Fixes bold text extracted as regular from PDFs whose embedded fonts carry no weight information other than the font header. (#3629)
• Fixed
  Fixes PDF form field iteration returning base-class objects instead of specific field types, such as text fields and checkboxes. (#3534)
• Fixed
  Fixes page rendering failing on PDFs that embed CFF2 OpenType fonts; affected text now renders with a substitute font. (#3579)
• Fixed
  Fixes PDF 2.0 text stored as UTF-8 — such as annotation contents, form field values, and bookmarks — reading back garbled. (#3505)
• Fixed
  Fixes PDF text extraction for Type 1 fonts with a scaled font matrix and for Euro signs mapped to a wrong character. (#3686)
• Fixed
  Fixes PDF text extraction to use the actual text a PDF gives for glyphs and to return ligatures such as fl as letters. (J#GDPDF-1537)
• Fixed
  Fixes redaction region placement on PDF pages whose page box doesn’t start at the origin. (L#GDPIC-1658, L#GDPDF-1532)
• Fixed
  Fixes rotated text, such as vertical table headers and axis labels, missing from PDF layout extraction. (#3625)
• Fixed
  Fixes structured data extraction failing for schemas that use local $defs or $ref definitions. (#3580)
• Fixed
  Fixes symbolic fonts that name glyphs by Unicode extracting raw character codes instead of the characters. (#3625)
• Fixed
  Fixes text drawn with a negative font size in a half-turned text matrix being extracted mirrored and rendered with misplaced spacing. (J#GDPDF-1344)
• Fixed
  Fixes the PDF reducer clearing the producer metadata of the output document when no producer name is set. (#3436)
• Fixed
  Fixes title metadata preservation and annotation placement when editing PDF tags. (#3627)
• Fixed
  Fixes XFDF export failing on FIPS-enabled hosts when the PDF contains file attachment annotations. (#3535)

Templating

• Fixed
  Fixes line breaks and tabs being dropped from values substituted into Office document templates. (J#OFTAPI-27)
• Fixed
  Fixes Office template processing failing when a section wraps a paragraph loop inside a single table cell. (J#OFTAPI-25)

Vision

• Added
  Adds a dedicated connection setting for the form field detect-extract model, plus an optional fallback flow and connection for form field enrichment. (#3452)
• Added
  Adds a Markdown EmitFigurePlaceholders setting; the bare [Figure] picture placeholder is now off by default. (L#NAT-626)
• Added
  Adds a vlm_batches_failed event to VLM-enhanced recognition that reports how many vision-model batches failed on a page and the first error, so a page that lost part of its text to a provider failure is visible in the logs instead of passing silently. (#3448)
• Added
  Adds an opt-in enableInlineFormatting option to DWS.Parse for bold, italic, and other inline styles in Markdown. (L#NAT-514)
• Added
  Adds an opt-in setting that reads a list printed in side-by-side column groups with a repeated header as one table. (L#NAVI-187)
• Added
  Adds description, value, and group ID to each detected form field in the JSON output, and reports the enrichment strategy used to produce each form’s fields. (#3452)
• Added
  Adds document-settings configuration for in-process parsing and schema-based extraction through DataExtraction. (#3424, #3608)
• Added
  Adds orientation correction to the Structure engine for rotated pages that require OCR. (#3555)
• Added
  Adds streamCandidateLabels to document splitting: per-page and per-segment document types, with a split at each type change. (#3649)
• Added
  Adds support for reading USPS Intelligent Mail barcodes in parsed documents, such as the barcode printed on an envelope or mail label. (L#DWS-869)
• Added
  Adds the font class to text from born-digital PDFs and cleans up their font family names in Vision output. (#3751)
• Added
  Adds the font family, class, and size of each paragraph and table cell to the Vision JSON output. (#3717)
• Changed
  Improves Markdown export by rejoining words hyphenated across line breaks using language-specific dictionaries. Disable joinHyphenatedLineBreaks to preserve those breaks. (L#NAT-627)
• Changed
  Improves Excel-to-Markdown conversion by recognizing each table, heading, and paragraph on a sheet. (#3528)
• Changed
  Improves checkbox and option-state extraction with VLM-enhanced recognition: Checkboxes, radio buttons, and tick boxes are transcribed in place as marked or empty boxes and are no longer lost when OCR and VLM text are merged. (#3447)
• Changed
  Improves OCR accuracy on scanned pages. (L#NAVI-130)
• Changed
  Improves OCR accuracy when the recognition pipeline runs directly: It now reads scans as the extraction modes do. (#3607)
• Changed
  Improves the reading order of scanned forms and receipts: A page the layout detector splits into many short one-line regions is now read row by row, which keeps each label next to its value. (#3450)
• Changed
  Improves VLM-enhanced recognition accuracy. (L#NAVI-124)
• Changed
  Improves word accuracy in VLM-enhanced recognition. (L#NAVI-125)
• Changed
  Improves AdaptiveOCR table output: rejects layout boxes misread as tables and emits enumerated items as lists. (#3520)
• Changed
  Improves AI data extraction from charts and diagrams with OpenAI models: The images of the pages that hold a chart or a picture are now sent to the model with their text by default. Set includePageImages to send every page. (#3782)
• Changed
  Improves AI data extraction of long lists on large documents by filling them from a second pass over smaller page windows. (#3778)
• Changed
  Improves AI extraction from very long or wide pages, such as well logs and time-location charts, with gpt-6 on OpenAI: Such a strip page is now sent as overlapping image tiles near its rendered resolution instead of one image the provider shrinks. (#3803)
• Changed
  Improves automatic document parsing speed by using available GPUs for page routing. (L#DWS-967)
• Changed
  Improves bold and italic detection on scanned pages, and adds font family, class, and size to recognized text. (#3717)
• Changed
  Improves chart data extraction with fast geometry-based bar chart reconstruction. (#3316)
• Changed
  Improves computed values in AI data extraction, such as totals, counts, and shares: A second pass works out the number fields the model left at 0, and OpenAI reasoning models now default to a high reasoning effort. (#3783)
• Changed
  Improves form field detection and fill-state accuracy with an updated detector and a dedicated fill-state model. (#3330)
• Changed
  Improves ICR table extraction for tables without printed grid lines, cut across page tiles, or read as text or form. (#3625)
• Changed
  Improves ICR table headers, columns, and splits, and keeps grouped header rows in Markdown table export. (#3625)
• Changed
  Improves table detection in VLM-enhanced ICR by asking the VLM whether grid-like text regions are tables. (#3625)
• Changed
  Improves table extraction on scanned documents. (L#NAVI-106, L#NAVI-107)
• Changed
  Improves table text in VLM-enhanced recognition. (L#NAVI-110)
• Changed
  Improves text recognition in tables, including tables pasted as images and text on shaded cells. (#3610)
• Changed
  Improves the reliability of model downloads. (L#NAVI-117)
• Changed
  Improves the text of machine-readable zones on passports, identity cards, and visas, and flags unconfirmed readings. (L#NAVI-189)
• Changed
  Changes the default OCR language to automatic detection, so each document is read in its own language. (L#NAVI-115)
• Changed
  Exports the charts embedded in an Excel workbook as data tables when converting it to Markdown or IR Lite. (#3826)
• Changed
  Markdown export now applies the lexicon pack of the configured OCR language (OcrSettings.DefaultLanguages) to born-digital text, whose words carry no language of their own, and OcrSettings.DetectLanguageFromDigitalText (advanced, off by default) identifies born-digital pages from their text without an OCR pass when an explicit language is configured. (L#NAT-633)
• Changed
  Changes layout reconstruction to require the VLM-enhanced ICR license, as the VLM setting does. (#3607)
• Changed
  Changes VLM-enhanced recognition to keep OCR running by default, so text outside the detected regions and table content is no longer missing from scanned pages. Where the VLM’s transcription of a region is kept, the region’s words spell that transcription; a word OCR didn’t find gets an estimated position, marked boundsEstimated in the JSON output. Set the OCR activation mode to deactivated to make the VLM the only text source. (#3446)
• Changed
  Changes document classification and splitting to read page text with AdaptiveOcr unless a vision engine is set. (#3809)
• Changed
  Changes the Vision.Decide context window from 8K to 32K tokens, which uses about 0.6 GB more memory per process. (#3698)
• Changed
  Changes Vision.Decide to answer requests it can’t accept with a 400 and code invalid_input instead of a 500. (#3728)
• Changed
  Changes Vision.Decide to fail when the decision model calibration is unavailable instead of returning uncalibrated probabilities. (#3708)
• Fixed
  Fixes layout reconstruction not running on scanned pages, now with the local vision model too. (#3607)
• Fixed
  Fixes lower OCR accuracy on GPU machines when several pages or documents share the GPU. A time limit canceled the GPU text-recognition pass while it waited for the GPU, so those pages lost that pass. The limit is now off by default, and a recognition pass that is cut short or fails is reported. (L#NAVI-127)
• Fixed
  Fixes words read twice on scanned pages when the recognition pipeline runs directly on a GPU machine. (#3607)
• Fixed
  Fixes a repaired word in a PDF text layer picking up text from the line above or below it. (#3620)
• Fixed
  Fixes a value printed beside its label being read before the label. (#3609)
• Fixed
  Fixes Auto-mode extraction failing with an internal error on documents whose pages all route to the structure tier. (#3541)
• Fixed
  Fixes Auto-mode pages routed to the understand and agentic tiers returning corrupted text on GPU machines: They now run the same evidence-fusion OCR pipeline the explicit modes use. (#3562)
• Fixed
  Fixes barcode blocks in parsed documents covering the whole surrounding picture instead of only the barcode. (L#DWS-853)
• Fixed
  Fixes document decision classification falling back to CPU when it’s the first operation on a GPU host. (#3800)
• Fixed
  Fixes duplicate text recognition when printed and handwritten regions partially overlap. (L#NAVI-69)
• Fixed
  Fixes extraction losing the text of a zone printed on the same line as another, such as a value beside its label. (#3609)
• Fixed
  Fixes form values and check marks that no detected region claimed being output in one block, away from their labels; each now follows the text beside or above it in the reading order. (#3611)
• Fixed
  Fixes grounding scores missing from the confidence results of windowed and schema-split extractions. (#3647)
• Fixed
  Fixes invoice headers whose multiline addresses were split into several table rows. (L#NAVI-134)
• Fixed
  Fixes label and value lines printed above a table being merged into the table. (L#NAVI-180)
• Fixed
  Fixes Markdown export fusing a hyphenated compound whose own hyphen sits at a line wrap, so quasi-experimental no longer becomes quasiexperimental. (L#NAT-625)
• Fixed
  Fixes missing font and bold or italic styles on words that Vision recognizes late on scanned pages, such as table cells, and missing font sizes on scanned images that declare their resolution. (#3756)
• Fixed
  Fixes ordinary words being reported as superscript or subscript on gray and low-contrast scans. (#3449)
• Fixed
  Fixes paragraphs that share a reading position appearing in a different order from run to run in the text layer of OCR’d PDFs. (#3616)
• Fixed
  Fixes reading order on pages where a value is printed beside its label, such as an email header or a totals box: Each label now reads with its value before the text below it. (#3621)
• Fixed
  Fixes scanned zones losing their text when the page’s language couldn’t be detected. (#3618)
• Fixed
  Fixes small tables with one or two rows of values being exported as text instead of a table. (L#NAVI-177)
• Fixed
  Fixes some QR codes and barcodes being returned as pictures instead of their decoded value. (#3475)
• Fixed
  Fixes Specialist form field VLM calls using best-effort (not strict) structured output, which could silently degrade labeling accuracy by letting malformed fields through instead of being rejected up front. (#3652)
• Fixed
  Fixes superscript formatting in Markdown exports when inline formatting is enabled. (L#NAVI-195)
• Fixed
  Fixes table cells that held the values of two neighboring columns. (L#NAVI-178)
• Fixed
  Fixes table cells that wrap over two lines being split into two rows. (L#NAVI-134)
• Fixed
  Fixes table header rows being exported as a heading above the table. (L#NAVI-179)
• Fixed
  Fixes tables whose first columns were exported as text beside the rest of the table. (L#NAVI-180)
• Fixed
  Fixes tables whose total line was merged into the last row. (L#NAVI-176)
• Fixed
  Fixes text extraction on rotated PDF pages. (L#NAVI-111)
• Fixed
  Fixes text painted in the same color as its background appearing in ICR output. (#3625)
• Fixed
  Fixes text printed twice when one recognized word covered both a label and its value, such as a tax number after its label. (#3617)
• Fixed
  Fixes the Specialist form field VLM creating fields not grounded in any real detection on the page. (#3669)
• Fixed
  Fixes checkbox and radio button states being misread and requires non-empty field labels and descriptions during AI form field extraction. (#3652)
• Fixed
  Fixes Vision coordinates that could differ slightly between two GPU runs on the same document. (L#NAVI-192)
• Fixed
  Fixes Vision extraction results that could change when the same document is processed again. (L#NAVI-192)
• Fixed
  Fixes Vision warmup skipping the local vision model that the Auto engine and layout reconstruction use. (#3614)
• Fixed
  Fixes Vision.Decide answering a question with empty instructions with a 500 instead of a 400 invalid_input. (#3753)
• Fixed
  Fixes words of two text lines read interleaved, and in a different order from run to run, when a tall graphic on the same line was recognized as text. (#3615)

General

• Added
  Adds an optional Fast mode for lower-latency OpenAI processing.
• Added
  Adds zero-configuration local vision processing with automatic model downloads and GPU acceleration. (J#NAT-551)
• Changed
  Improves born-digital text extraction by detecting citation markers, footnote numbers, and bracketed references as super/subscript and promoting heading candidates found in the text layer to heading zones.
• Changed
  Improves the agentic vision tier by reporting a structural role for each text zone and promoting high-confidence Title and SectionHeader detector Text zones to headings.
• Changed
  Improves vision layout on born-digital pages by separating larger or bold heading lines inside detected text zones into their own heading zones. This behavior is controlled by splitHeadingsInTextZones, which is on by default.
• Changed
  Improves chart data extraction on multipanel and multilevel charts by asking the model for one table per panel, including the panel title in a <caption>, and folding extra grouping levels into the row label.
• Changed
  Changes the default OpenAI vision model to gpt-5.6-luna.
• Fixed
  Fixes an issue where vision extraction failed with Failed to crop image to zone bounds on born-digital pages whose ruled frame extended past the page box. It also prevents decorative horizontal rules from being admitted as page-wide tables.
• Fixed
  Fixes an issue where License.TraceFeatures returned an empty or single-entry list in trial mode. It now reports every licensed component the application used.

OCR

• Added
  Adds form field fill-state detection so detected form fields now report whether they’re filled or empty, including checked/unchecked checkboxes, signed/blank signatures, and filled/empty text fields.
• Added
  Adds checkbox and form field state to Markdown export. Checkboxes render as task-list items, and labeled form fields render as label: value lines.
• Added
  Adds the Ocr.detectLanguageOncePerDocument setting, which controls whether automatic language detection runs once for the whole document or independently on every page.
• Added
  Adds a language tag to Markdown code fences, inferred from the code block’s own content.
• Added
  Adds document splitting, which finds sub-document boundaries in a merged document and returns the page ranges.
• Changed
  Improves Markdown export of text that wraps across lines. A word broken by a hyphen at the end of a line is rejoined into one word, while a wrapped number range keeps its hyphen.
• Changed
  Speeds up automatic OCR language detection by identifying the language from a band of the page instead of the whole page. This behavior is controlled by the new Ocr.detectLanguagePageFraction setting.
• Changed
  Speeds up automatic OCR language detection by running its text-recognition pass for speed rather than accuracy. This behavior is controlled by the new Ocr.detectLanguageFavorAccuracy setting.
• Changed
  Changes zero-shot classification so single-label and multilabel modes may now return different predicted labels and rankings.
• Changed
  Changes table export so a header row detected by the table model is emitted with header cells instead of plain cells.
• Changed
  Changes the classification confidence threshold to apply on each mode’s own confidence scale, so single-label and multilabel documents may abstain differently.
• Changed
  Changes Markdown export to apply inline bold and italic styling to headings and page headers and footers, and to emit runs that are both bold and italic as **_text_** rather than ***text***.
• Changed
  Changes Markdown export to include page headers and footers by default, fenced by page-header and page-footer comment markers.
• Fixed
  Fixes an issue where Markdown export dropped every chart table after the first when a chart held several panels.
• Fixed
  Fixes an issue where table cell column and row spans reached past the end of the table, which produced malformed tables in the Markdown and JSON exports.
• Fixed
  Fixes an issue where table extraction on tall tables whose recognized structure collapsed to a couple of rows lost rows. The table region is now reread in horizontal bands, and the lost rows are recovered.

PDF

• Added
  Adds bold detection from font weight metadata in born-digital PDF text extraction.
• Changed
  Improves text style extraction by detecting underlines in born-digital PDFs and resolving bold and italic styles from font descriptors.

Vision

• Added
  Adds a reasoning-effort setting for AI data extraction and vision zone classification, trading model reasoning time against cost.
• Changed
  Improves chart detection when extracting document content with vision language models.
• Changed
  Improves NLI grounding confidence on numeric table cells by anchoring the verbalized claim to the cell’s printed value and identifying its row by all its columns, so correct numeric extractions no longer receive low groundedness.
• Changed
  Changes AI data extraction to enforce the requested JSON schema on the provider and fall back automatically for models that don’t support it.

General

• Changed
  Improves exception-to-status mappings and adds DocumentConversionFailed and VisionOperationFailed status codes. (J#NAT-603)
• Changed
  Improves first-request latency on GPU deployments by preparing the CUDA runtime during warmup instead of during the first request.
• Changed
  Improves multipage CPU processing throughput by scheduling concurrent page inference on a shared thread pool.
• Fixed
  Fixes GPU memory growing across requests and never being released when vision inference runs on the CUDA provider.

OCR

• Added
  Adds a maximum-threads OCR setting that processes pages concurrently when making a document searchable.
• Added
  Adds offline language detection for Tamil, Telugu, Kannada, Malayalam, and Bengali, with faster script models.
• Changed
  Improves OCR speed on modern x86 processors by using VNNI instructions when available.
• Changed
  Optimizes automatic OCR language detection: detects once per document and skips it when page OCR is disabled.
• Fixed
  Fixes the OCR accuracy-vs.-speed setting being ignored. Disabling accuracy mode now selects the faster recognition path.

PDF

• Added
  Adds a PDF validation API that verifies PDF/A and PDF/UA conformance and produces detailed validation reports.
• Added
  Adds dedicated OCR and PDF editing exception types that identify the failed operation and page.
• Added
  Adds the AutoTaggingEngine accessibility setting to choose between Heuristic, Vision, and VisionVlm tagging.
• Added
  Adds vision-backed layout detection to PDF/UA auto-tagging for higher-fidelity document structure.

PDF

• Fixed
  Fixes MakeSearchable writing every page’s OCR text layer onto the first page of multipage documents.

General

• Added
  Adds GPU-accelerated vision inference on Linux via the WebGPU execution provider on Vulkan-capable GPUs. (J#NAT-560)
• Added
  Adds Windows GPU-accelerated vision inference via the WebGPU execution provider on Direct3D 12. (J#NAT-560)
• Added
  Adds a local in-process vision language model provider for AI document augmentation without a cloud or server dependency. (J#NAT-521)
• Added
  Adds a natural language inference (NLI) groundedness confidence and an opt-in v2 result.json metadata schema for AI document extraction.
• Added
  Adds optional document restoration that enhances extraction for perspective-distorted document photos before extraction. (#2863)
• Added
  Adds support for newer Anthropic models in the Vision API. (J#NAVI-32)
• Added
  Adds support for opening password-protected PDF documents by supplying the password through the document open settings.
• Changed
  Renames Claude vision provider APIs to Anthropic. (J#NAVI-31)
• Fixed
  Fixes a process-exit crash when GPU vision inference is requested on Linux hosts without a Vulkan-capable GPU. (J#NAT-579)
• Fixed
  Fixes document classification reporting only one processed page for multipage documents.
• Fixed
  Fixes error handling to throw proper exceptions for file-not-found and invalid handle states.
• Fixed
  Fixes nullable numeric getters returning errors for unset values in the Java and Python SDKs. (#NATL-586)

OCR

• Fixed
  Fixes AdaptiveOcr extraction failing on PDF pages whose embedded text layer contains no extractable words.

PDF

• Added
  Adds page-level automatic deskewing to the PDF editor.

Templating

• Added
  Adds PowerPoint and Excel template-to-PDF/UA generation via new presentation and spreadsheet editors. (J#NAT-559)

General

• Added
  Adds Vision.GenerateSchema for generating extraction-ready JSON Schemas with optional cross-field constraint rules.
• Added
  Adds a diagnostics artifact to vision document extraction containing VLM token usage and provider context when includeDiagnostics is enabled. (J#NAT-501)
• Added
  Adds automatic orientation correction that straightens pages rotated 90/180/270 degrees before vision extraction. (J#NAVI-18)
• Added
  Adds barcode recognition for 1D, QR, Data Matrix, PDF417, Aztec, Micro QR, and MaxiCode codes in the vision pipeline. (J#NAT-530)
• Added
  Adds page selection to document extraction through a 1-based page-range setting such as 1–3,5. (J#NAT-464)
• Added
  Adds quad-point polygon geometry to document regions for accurate representation of rotated and skewed areas. (J#NAVI-23)
• Added
  Adds ranked BM-25 text search over documents: Build a reusable index, and then query it for passages by page and line. (J#NAT-512)
• Changed
  Improves formula recognition performance, roughly halving processing time on equation-heavy documents. (J#NAT-556)
• Changed
  Improves table extraction speed by up to 6× on table-heavy documents. (J#NAT-524)
• Changed
  Changes vision licensing checks to match exactly the capabilities each API call invokes.
• Changed
  Disables local handwriting OCR by default, with an advanced setting to opt back in. (#2746)
• Changed
  Optimizes vision document extraction with GPU acceleration on macOS, up to ~3× faster. (J#NAT-524, J#NAT-525, J#NAT-537, J#NAT-550, J#NAT-552, J#NAT-554, J#NAT-555, J#NAT-556)
• Changed
  Updates the default OpenAI and Anthropic vision models to gpt-5.4 and claude-sonnet-4-6. (#2964)
• Fixed
  Fixes AI extraction errors mislabeled as an unreachable endpoint when the model returns no content.
• Fixed
  Fixes Markdown table-cell export replacing en/em dashes and minus signs with hyphens and leaving formatting noise.
• Fixed
  Fixes a crash when processing documents concurrently with GPU-accelerated vision. (J#NAT-556)
• Fixed
  Fixes a memory leak where vision objects weren’t released after use, growing native memory over time. (J#NAT-515)
• Fixed
  Fixes an issue where the structured JSON export flattened multiline zones into a single line; zone text now preserves visual line breaks as newlines. (#2771)
• Fixed
  Fixes duplicate paragraph text appearing alongside extracted table content when table extraction is enabled. (#2953)
• Fixed
  Fixes token usage counts to include Gemini thinking tokens and OpenAI cached input tokens.

OCR

• Added
  Adds fully offline language and text-direction detection for documents, with a text-only mode that needs no document.
• Added
  Adds multilingual zero-shot document classification, automatically handling non-English documents and labels.
• Added
  Adds text-only zero-shot classification for text you already have, with no document opened or rendered.
• Added
  Adds zero-shot document classification scoring text and image against caller-supplied labels.
• Changed
  Improves concurrent processing of vision documents. (J#NAVI-21)
• Changed
  Improves offline language detection by suppressing spurious secondary languages via new OCR settings.
• Fixed
  Fixes a race condition that could corrupt results when running vision OCR extraction concurrently. (J#NAVI-19)
• Fixed
  Fixes data tables being misclassified as a table of contents and dropped during vision data extraction.

General

• Changed
  Improves classification for handwriting zones for the VLM pipeline. (#2737)
• Changed
  Improves the Python API reference layout with per-tag folder hierarchies for settings classes. (J#NAT-481)
• Changed
  Changes the OCR setting to support multiple modes, including an automatic mode that selects the best processing pipeline. (#2738)

Office

• Added
  Adds rendering of DisplayBarCode QR code fields in Word documents during conversion. (J#MAMBO-2058)
• Added
  Adds rendering of form checkboxes and dropdowns from ODT documents during conversion. (J#MAMBO-2072)
• Fixed
  Fixes a crash when converting Word documents containing an IF mail-merge field that resolves to an empty result. (J#MAMBO-2076)
• Fixed
  Fixes DOCX repair so footers stored with an invalid hdr root element are corrected to ftr. (J#MAMBO-2078)
• Fixed
  Fixes an invalid image stream exception. (J#MAMBO-2077)
• Fixed
  Fixes MSG files with custom IPM.Note subclasses, such as archived emails, being reported as an unknown format. (J#GDPIC-1621)

PDF

• Changed
  Improves embedded font subsetting with corrected .notdef glyph handling and CFF font reconstruction.
• Changed
  Changes the Producer and Creator metadata of generated PDF documents from GdPicture.NET to Nutrient SDK. (J#NAT-460)
• Fixed
  Fixes PDF/A validation errors related to embedded font glyph and width consistency. (J#PDFA-95)

General

• Added
  Adds a text export format for converting documents to plain text.
• Added
  Adds an opt-in strict structured-output mode for AI data extraction with automatic schema normalization.
• Added
  Adds first-class form field detection to the Vision API, with optional AI semantic labeling of detected fields.
• Added
  Adds intelligent chart parsing via HTML tables. (J#NAT-457)
• Added
  Adds PNG, JPEG, and BMP support to image export, with automatic format detection from the destination filename. (J#NAT-474)
• Added
  Adds single-flight coordination in ResourceDownloader to coalesce concurrent requests for the same model download. (J#NAT-472)
• Added
  Adds support for primitive, enum, and string-array properties in settings. (J#NAT-107)
• Added
  Adds Vision.ExtractStructured for AI-powered structured data extraction shaped by a caller-supplied JSON Schema.
• Changed
  Improves AI extraction robustness with fail-fast schema validation and culture-independent confidence scoring.
• Changed
  Improves AI model download reliability with integrity validation and automatic retry on failure. (J#NAT-419)
• Changed
  Improves form field detection by removing overlapping predictions for cleaner, more consistent results. (J#NAT-471)
• Changed
  Improves heuristic PDF layout extraction.
• Changed
  Improves the Office-to-PDF speed tests. (J#MAMBO-2073)
• Removed
  Removes form field inference from vision intelligent content recognition. (J#NAT-469)
• Fixed
  Fixes intermittent errors in the VlmEnhanced vision pipeline when processing documents concurrently. (J#NAVI-16)
• Fixed
  Fixes the GdPicture.NET.Toolkit NuGet package on .NET Framework 4.6.2 by including native runtimes as a package dependency. (J#GDPIC-1619)
• Fixed
  Fixes incorrect text appearing inside images in multipage document Markdown output. (#2654)
• Fixed
  Fixes loss of image alternate text descriptions when converting documents to PDF/UA. (J#PDFA-93)
• Fixed
  Fixes word alignment in document layout that caused paragraphs to absorb text from neighboring lines. (#2633)

Office

• Changed
  Improves DOCX line spacing rendering for documents using the legacy line spacing compatibility option. (J#MAMBO-2027)
• Changed
  Improves ODT-to-PDF page dimension rendering to match Microsoft Word. (J#MAMBO-2059)
• Changed
  Improves parsing of empty spaces from ODT documents. (J#MAMBO-2075)
• Fixed
  Fixes an exception when loading a Word document that contains an empty custom XML part. (J#MAMBO-2067)
• Fixed
  Fixes an exception when rendering table cells where all text falls outside the cell boundary. (J#MAMBO-2065)
• Fixed
  Fixes missing spacing before a numbered paragraph when the preceding paragraph isn’t part of the list. (J#MAMBO-2062)
• Fixed
  Fixes ODT tables losing their header row. (J#MAMBO-2074)

PDF

• Added
  Adds support for converting between Instant JSON and GdPicture XMP annotations. (J#NAT-406)
• Added
  Adds support for exporting Instant JSON. (J#NAT-397)
• Changed
  Improves PDF-to-Markdown conversion performance.
• Changed
  Enables change tracking for Instant JSON exports by default. (J#NAT-394)
• Fixed
  Fixes empty embedded file handling when merging PDF documents. (J#GDPDF-1520)

Templating

• Added
  Adds PowerPoint (PPTX) templating with placeholders, loops, conditionals, tables, and images to the Office templater. (J#OFTAPI-20)

• Added
  Adds the AdaptiveOcr vision engine — a heuristic-first PDF extraction pipeline with automatic OCR fallback for image-based pages.
• Added
  Adds a born-digital text fast path to VLM pipelines and enables VLM-only text output when OCR is disabled. (#2607)
• Added
  Adds direct PDF processing support to the Vision API for improved document analysis.
• Added
  Adds support for creating searchable PDFs.
• Added
  Adds form detection support.
• Changed
  Improves licensing error information when Document.open fails because the input format requires a feature not enabled by the active license. (J#NAPY-4)
• Changed
  Improves vision data extraction fidelity.
• Fixed
  Fixes ARM64 Windows compatibility issues where the SDK failed with cryptic errors on ARM64 Windows machines. (J#NAPY-6)
• Fixed
  Fixes caption and footnote relationships being linked to incorrect element IDs. (#2615)
• Fixed
  Fixes Document.open to raise FileNotFoundException for missing files. (J#NAPY-1)
• Fixed
  Fixes a crash when running OCR on Linux. (J#NAT-458)
• Fixed
  Fixes the SDK version displayed in the license banner to match the installed wheel version. (J#NAPY-3)
• Changed
  Improves Python SDK conversion, rendering, and extraction capabilities by incorporating enhancements from the Nutrient .NET SDK 14.4.3(opens in a new tab) release line.

• Added
  Adds Markdown export format to Vision API content extraction, including options for semantic block formatting and header/footer inclusion. (J#NAT-347 and #2590)
• Added
  Adds a failure message when VlmEnhancedIcr isn’t configured correctly. (#2534)
• Changed
  Improves SdkSettings accessors with property-style syntax and refines the Python API reference documentation. (J#NAT-441)
• Changed
  Improves Vision API JSON export performance on large multipage documents. (#2499)
• Changed
  Improves Python SDK conversion, rendering, and extraction capabilities by incorporating enhancements from the Nutrient .NET SDK 14.4.2(opens in a new tab) release line.

• Added
  Adds compatibility with Python >=3.8.
• Changed
  Improves Python SDK conversion, rendering, and extraction capabilities by incorporating enhancements from the Nutrient .NET SDK 14.4.1(opens in a new tab) release line.
• Changed
  Improves handwriting vision capabilities by adding a new in-house model that better identifies handwritten regions in images.

• Changed
  Improves Python error handling and exception mapping for clearer, more specific runtime errors. (#2383)
• Changed
  Improves stability and reliability in generated Python bindings. (#2383)

• Fixed
  Fixes XLSX conversion and OCR memory usage issues as part of a coordinated .NET/Java/Python SDK release. (#2309)
• Changed
  Improves the Vision API for intelligent content recognition (ICR) and fixes issues in the vision language model (VLM) pipeline. (#2309)
• Changed
  Improves macOS distribution compatibility for package delivery. (#2309)

• Added
  Adds the first public Python SDK release, with complete feature parity with Nutrient Java SDK(opens in a new tab) (including Vision) and near parity with Nutrient .NET SDK(opens in a new tab). (#2211, #2221)
• Changed
  Positions the SDK for rapid prototyping and as an entry point for complex document manipulation workflows. (#2211, #2221)
• Changed
  Highlights new Python guides(opens in a new tab) and the newly published Python API reference(opens in a new tab) as part of the release. (#2211, #2221)