docling

Feature

• Create a backend parser for XBRL instance reports (#3017) (334ba6e)
• Unified model-family inference engines (including image-classification) and KServe v2 API support (#2979) (0353293)

Fix

• Skip ASR segments when length is zero (#2998) (6b824f8)
• docx: Guard against None hyperlink address in _get_paragraph_elements (#2367) (#3022) (236216e)

Feature

• Introduce docling-parse v5 and deprecate old docling-parse backends (#2872) (bf417e6)

Fix

• Security vulnerabilities with XML External Entity and related attacks (#3009) (576bada)
• csv: Set default delimiter by default (#3005) (a1b0e3f)
• Improved deserialization of engine_options (#3008) (dbba6ea)

Fix

• asciidoc: Handle commas in image alt text (#2983) (86b6912)
• Use timezone-aware datetime (#2947) (e2870f9)
• Add failed pages to DoclingDocument for page break consistency (#2939) (1f91482)

Feature

• Inference engines abstraction for object detection model family with HF Transformers and ONNX runtime (#2959) (14e474c)
• Added support for parsing LaTeX (.tex) documents (#2890) (e6ccb8b)
• Introduce pluggable VLM runtime system with preset-based configuration (#2919) (d4c8713)

Fix

• Restore expected behavior for artifacts_path and accelerator_options in VLM engines (#2961) (9721321)
• Allow offline chart extraction model artifacts (#2957) (ae4fdbb)

Documentation

• Add LaTeX and WebVTT as supported types (#2974) (704ef0a)

Feature

• Add chart extraction models (#2848) (fe45c71)

Fix

• backend: Improve Excel table bounds detection and flatten merged cells (#2778) (3110c43)
• pptx: Handle picture shapes with external image references (#2914) (5e452a2)

Documentation

• Add granite vision for charts (#2946) (a5ad8f2)

Feature

• Webvtt and source tracker (#2787) (0602a7c)
• Add support for Word document comments extraction (#2834) (b6ca094)

Fix

• Allow newer typer versions (#2930) (6f205ae)
• rapidocr: Use new model links for RapidOCR (#2928) (82b7982)
• Presets for ollama (#2926) (4a269de)

Feature

• Drop support for Python 3.9 (#2905) (7f38658)

Fix

• md: Handle pipe symbols that are not table markers (#2904) (86eaef5)
• Remove direct vllm dependency (#2910) (7a1952a)
• PPTX parsing: bullet points not grouped correctly under subheadings (#2663) (#2855) (999dbb2)

Documentation

• Add comprehensive docstrings to PdfPipelineOptions (#2827) (ab91786)

Fix

• Off-by-one error for page indexing in vlm_pipeline (#2902) (08f49e2)

Feature

• New picture classifier v2.0 (#2889) (43badc3)
• Add classification filters for picture description (#2836) (ac16a26)

Fix

• Torch compatibility for xpu (#2894) (00273f6)
• Standardize page_no to 1-based indexing (#2847) (1b4d82d)
• Usage of direct logging (#2884) (2fe9def)
• Relax pypdfium2 version constraint, support 5.x (#2873) (daf2bc6)

Documentation

• Correct broken link to supported formats (#2878) (16e88d5)

Feature

• Support for DeepSeek-OCR in VLM pipeline (#2798) (19af03f)

Fix

• logging: Include page numbers in preprocess error messages (#2858) (89bea24)
• docx: Handle grouped pictures (#2861) (5c1f8f0)

Documentation

• Fix Colab badge links and Weaviate typo in docs examples (#2871) (72851cc)
• example: Fix update sample image path to be relative (#2864) (211c759)
• Add Semantica integration (#2860) (bf80e32)

Feature

• Enrichment annotations in the new meta format (#2859) (aab3ff5)
• Add XPU device support for Intel GPUs (#2809) (2b83fdd)
• Add option to report timings details (#2772) (cbc6537)

Fix

• Lock new deps and update python 3.14 warnings (#2844) (d9295df)
• Correct type hint for table_structure_options usage (#2823) (a0530a2)
• Transformers models lazy-loaded (#2826) (3ef4525)
• Font download by passing font_path to RapidOcr (#2822) (ffafe58)
• cli: Add Layout and Table models to --show-external-plugins (#2832) (ed57089)

Feature

• Add preset for using granite-docling via vllm and other apis (#2792) (241d19e)

Fix

• docx: Handle tables with merged cells causing IndexError (#2813) (faff935)
• markdown: Allow text before headers also in mixed markdown and html (#2801) (595115d)

Documentation

• RTX: Guidelines for best performance on RTX GPUs (#2765) (be085c0)
• Add docstrings to DocumentConverter #2748 (#2782) (cc5e3ce)
• style: Fix link visibility in dark mode (#2804) (150fe90)

Feature

• Add YAML output format to CLI (#2768) (da7678a)

Fix

• rapidocr: Use correct parameter name for rec_keys_path (#2762) (1d78418)
• docx: Handle missing value in paragraph style name (#2761) (a97d950)

Documentation

• Add Pydantic field documentation for PipelineOptions (#2771) (7c24b01)
• gpu: Add benchmarks of standard pipeline with OCR (#2764) (d03439c)

Fix

• Clear word/char cells when force_full_page_ocr is used (#2738) (1df0560)
• Add missing font download in the rapidocr artifacts (#2735) (edbabfc)
• Ensure proper image_scale for generated page images in VLM pipelines (#2728) (609069d)
• html: Tackle paragraphs with block-level elements (#2720) (d007ba0)
• html: Prevent hierarchy reset in rich table cells (#2716) (aebe25c)
• docx: Parse integrals as n-ary objects without chr element (#2712) (c97715f)

Feature

• experimental: Add experimental TableCropsLayoutModel (#2669) (1344362)
• Factory and plugin-capability for Layout and Table models (#2637) (ad97e52)

Fix

• InputFormat.IMAGE must have correct pipeline (#2707) (6ef4ffd)
• Do not consider singleton cells in xlsx as TableItems but rather TextItems (#2589) (54cd6d7)
• docx: Missing list items after numbered header (#2665) (e580554)

Documentation

• Example on how to apply external OCR as post processing (#2517) (fa21128)
• More GPU results and improvements in the example docs (#2674) (b75c646)
• Fix typo on jobkit page (#2671) (146b4f0)

Feature

• Add save and load for conversion result (#2648) (b559813)

Fix

• Respect document_timeout in new threaded StandardPdfPipeline (#2653) (2087c6b)
• In DocumentConverter.convert_string() make nullable name parameter optional (#2660) (6fb9a5f)
• Enable GPU for RapidOCR when available (#2659) (463a3fd)
• Remove py3.14 requirement for default rapidocr (#2639) (da4c2e9)

Documentation

• Add Hector as compatible AI agent platform integration (#2662) (ce5a099)
• Added documentation to use SuryaOCR via plugin docling-surya (#2533) (b216ad8)
• Fix broken homepage links (#2651) (03e7c7d)
• examples: Processing parquet file of images (#2641) (8af228f)
• Move Installation and Quickstart (Usage) under Getting started (#2644) (d549445)
• Add redirection from getting started page (#2640) (ac9fc58)
• examples: Remove deprecation warnings with export_to_dataframe (#2638) (f552862)

Feature

• Add the Image backend (#2627) (3495b73)
• experimental: Layout + VLM model with layout prompt (#2244) (4852d8b)

Fix

• Correct the model-repo name (#2624) (14b436d)
• docx: Parse page headers and footers (#2599) (054c4a6)

Documentation

• Combine Home and Getting Started pages (#2600) (ae30373)

Fix

• Default to EasyOCR in Python 3.14 (#2605) (5c27567)

Fix

• docx: Slow table parsing (#2553) (ef623ff)
• html: Slow table parsing (#2582) (0ba8d5d)

Documentation

• Make navigation menus collapse and expand (#2573) (8da3d28)

Feature

• vlm: Track generated tokens and stop reasons for VLM models (#2543) (6a04e27)

Fix

• Temporarily pin NuExtract to working revision (#2588) (fa92574)
• ocr: Use PSM integer values directly instead of constructor (#2578) (1a5146a)

Fix

• Extract response from api_image_request in picture description (#2571) (8360aa5)

Feature

• Use threading in the standard pipeline and move old behavior to legacy (#2452) (268d027)

Fix

• pdf: Threadsafe for pypdfium2 backend (#2527) (a51275d)

Documentation

• Update link to Open WebUI docs (#2549) (01577e9)
• Update installation options with extras and review FAQ (#2548) (cb10043)
• Fix typos (#2546) (741c44f)

Feature

• vlm: Add num_tokens as attribtue for VlmPrediction (#2489) (b6c892b)
• Support for Python 3.14 (#2530) (cdffb47)

Fix

• Xlsx cell parsing, now returning values instead of formulas (#2520) (d9c90eb)

Documentation

• Add details and examples on optimal GPU setup (#2531) (97aa06b)
• Update opensearch notebook and backend documentation (#2519) (9a6fdf9)

Feature

• pdf: Support for password-protected PDF documents (#2499) (bbe82a6)
• backend: Add generic options support and HTML image handling modes (#2011) (a30e6a7)
• ASR: MLX Whisper Support for Apple Silicon (#2366) (657ce8b)

Fix

• markdown: Set the correct discriminator in md backend options (#2501) (4227fcc)
• xlsx: Speed up by detecting the true last non-empty row/column (#2404) (b66624b)

Documentation

• Fix typo in mcp.md (#2502) (86556d8)
• Discord badge with join link (#2473) (dd03b53)

Performance

• Use docling-parse-v4 as default (#2503) (89820d0)

Feature

• docx: Process drawingml objects in docx (#2453) (1682993)

Fix

• Use proper page concatentation in VLM pipeline MD/HTML conversion (#2458) (cd7f7ba)

Documentation

• Example on PII obfuscation (#2459) (3e6da2c)

Fix

• Avoid downloading easyocr models by default (#2454) (688a7df)

Feature

• AutoOCR model selecting the best OCR model available and deprecating the usage of EasyOCR (#2391) (f7244a4)
• Add Tesseract PSM options support (#2411) (f11f8c0)

Fix

• asr: Implement robust status check in AsrPipeline (#2442) (db985bb)
• Deal with chartsheets in workbooks (#2433) (cce18b2)
• Skip temporary docx files (#2413) (ee55013)
• AsrPipeline to handle absolute paths and BytesIO streams correctly (#2407) (b5f7fef)
• Enrichment of documents without pages metadata (pptx and xlsx) (#2401) (0610d01)
• Proper heading support in rich tables for HTML backend (#2394) (9705f40)

Documentation

• Remove deprecated call in custom_convert.py (#2447) (9020044)
• Fixed a few typos (#2441) (2a0f563)
• Add MongoDB + VoyageAI (#2382) (f2854b2)
• Add RAG example with MongoDB Atlas Vector Search and VoyageAI embeddings (#2341) (8a4b946)

Fix

• markdown: Setext heading support (#2359) (ee73ffa)
• docs: Fixed the color scheme (#2371) (246de77)
• Empty table handling (#2365) (ca2be7f)
• Add table raw content when no table structure model is used (#1815) (4f295ed)

Documentation

• Example using Hashicorp Vault PII transform (#2373) (a975a79)
• Jobkit and connectors (#2357) (e6c3b05)

Feature

• Repetition-based StoppingCriteria for GraniteDocling (#2323) (1e9dc43)
• Rich tables support for HTML backend (#2324) (c803abe)

Fix

• Pin wider range of typer (#2309) (68ae7cc)
• Update Transformers & VLLM inference code, CLI and VLM specs (#2322) (654c70f)
• Support escaped characters in markdown backend (#2304) (9d67bb9)

Documentation

• styling: Update color scheme (#2154) (325877a)
• vlm: Update SmolDocling to GraniteDocling references (#2315) (a873200)

Feature

• Rich tables for MSWord backend (#2291) (e2482a2)
• Add a backend parser for WebVTT files (#2288) (46efaae)

Fix

• Correct y-axis scaling in draw_table_cells (#2287) (b5628f1)

Documentation

• Update API VLM example with granite-docling (#2294) (8b7e83a)
• Fix examples rendering (#2281) (8322c2e)

Feature

• Add granite-docling model (#2272) (17afb66)
• RapidOcr: Support generic extra arguments for RapidOcr (#2266) (0e95171)

Fix

• Handle empty result from RapidOCR to avoid crash (#2264) (609d902)

Documentation

• Describe examples (#2262) (ff351fd)

Feature

• Enrichment steps on all convert pipelines (incl docx, html, etc) (#2251) (2c91234)

Fix

• Add missing features in ThreadedStandardPdfPipeline (#2252) (0700af2)
• Address deprecation warnings of dependencies (#2237) (c696549)

Documentation

• Add an example of RAG with OpenSearch (#2238) (f8cc545)
• Add instructions for using Docling with MCP to README (#2219) (e5cd702)
• Document VLM support requirement in extraction example (#2231) (55f5f37)

Feature

• Updating default parameters to get better performance with docling-parse (#2208) (b49d1ad)
• Updated the backend for new docling-parse (#2187) (b3d7542)

Documentation

• Add information extraction example (#2199) (a9f41b0)

Feature

• Heron layout model as new default (#1971) (e38aa0f)

Fix

• html: Access to variable not yet declared (#2171) (293e81b)

Feature

• [Beta] Extraction with schema (#2138) (9f4bc5b)
• msexcel: Set ContentLayer.INVISIBLE for invisible sheet (#1876) (a283ccf)

Fix

• pypdfium2: Fix OCR bounding box misalignment caused by mismatched rotation metadata (#2039) (4d94e38)
• Translation example (#2166) (9f0286b)
• Extend offline mode for rapidocr fonts (#2155) (9904d14)

Documentation

• Enrich landing pages (#2165) (96cab6b)

Feature

• Upgrade to RapidOCR 3.x (#2088) (3f60a0f)

Fix

• html: Preserve code blocks in list items (#2131) (fa3327e)

Fix

• Vllm extra only for linux x86_64 (#2126) (488f6cd)

Feature

• CLI: Option to download arbitrary HuggingFace model (#2123) (cdf079d)
• Batching support for VLMs in transformers backend, add initial VLLM backend (#2094) (3c660c0)
• html: Support formatting tags in HTML texts (#2111) (94fcc46)

Fix

• Improve numbered list detection for msword docs (#2100) (3f03709)

Documentation

• DPK pipeline example using docling library (#2112) (e76298c)
• Add Getting Started page (#2113) (8996d61)

Feature

• New code formula model (#2042) (d2494da)

Fix

• HTML: Parse footer tag as a group in furniture content layer (#2106) (c5f2e2f)

Performance

• Clean up resources with docling-parse v4, no parsed_page output by default (#2105) (5f57ff2)
• Speed up function _parse_orientation (#1934) (8820b55)

Feature

• Add backend for METS with Google Books profile (#1989) (31087f3)
• html: Support in-line anchor tags in HTML texts (#1659) (9687297)
• vlm: Ability to preprocess VLM response (#1907) (5f050f9)

Documentation

• Add docling Quarkus integration (#2083) (76c1fbd)

Feature

• Add convert_string to document-converter (#2069) (b09033c)

Fix

• html: Parse rawspan and colspan when they include non numerical values (#2048) (ed56f2d)
• Support new mlx-vlm module (#2001) (0130e3a)
• Extend error reporting when verbose logging is enabled (#2017) (2eb760d)
• HTML: Replace non-standard Unicode characters (#2006) (86f7012)

Documentation

• Add Langflow integration (#2068) (e2cca93)
• Add Arconia integration (#2061) (bfda6d3)

Feature

• Threaded PDF pipeline (#1951) (aed772a)

Fix

• markdown: Ensure correct parsing of nested lists (#1995) (aec29a7)
• HTML: Remove an unnecessary print command (#1988) (945721a)

Fix

• HTML: Concatenation of child strings in table cells and list items (#1981) (5132f06)
• docx: Adding plain latex equations to table cells (#1986) (0b83609)
• Preserve PARTIAL_SUCCESS status when document timeout hits (#1975) (98e2fcf)
• Multi-page image support (tiff) (#1928) (8d50a59)

Documentation

• Add chat with dosu (#1984) (7b5f860)

Fix

• Keep formula clusters also when empty (#1970) (67441ca)

Documentation

• Enrich existing DoclingDocument (#1969) (90a7cc4)
• Add documentation for confidence scores (#1912) (5d98bce)

Feature

• Add option to control empty clusters in layout postprocessing (#1940) (a436be7)

Fix

• Safe pipeline init, use device_map in transformers models (#1917) (cca05c4)
• Fix HTML table parser and JATS backend bugs (#1948) (e1e3053)
• KeyError: 'fPr' when processing latex fractions in DOCX files (#1926) (95e7096)
• Change granite vision model URL from preview to stable version (#1925) (c5fb353)

Documentation

• Fix typos (#1943) (d6d2dbe)

Feature

• Layout model specification and multiple choices (#1910) (2b8616d)
• Enable precision control in float serialization (#1914) (ec588df)
• Add image-text-to-text models in transformers (#1772) (a07ba86)
• vlm: Dynamic prompts (#1808) (b8813ee)

Fix

• ocr-utils: Unit test and fix the rotate_bounding_box function (#1897) (931eb55)
• Docs are missing osd packages for tesseract on RHEL (#1905) (e25873d)
• Use only backend for picture classifier (#1904) (edd4356)
• Typo in asr options (#1902) (dd8fde7)

Feature

• Introduce LayoutOptions to control layout postprocessing behaviour (#1870) (ec6cf6f)
• Integrate ListItemMarkerProcessor into document assembly (#1825) (56a0e10)

Fix

• Secure torch model inits with global locks (#1884) (598c9c5)
• Ensure that TesseractOcrModel does not crash in case OSD is not installed (#1866) (ae39a94)

Performance

• msexcel: _find_table_bounds use iter_rows/iter_cols instead of Worksheet.cell (#1875) (13865c0)
• Move expensive imports closer to usage (#1863) (3089cf2)

Feature

• Leverage new list modeling, capture default markers (#1856) (0533da1)

Fix

• markdown: Make parsing of rich table cells valid (#1821) (e79e4f0)

Fix

• Updated granite vision model version for picture description (#1852) (d337825)
• markdown: Fix single-formatted headings & list items (#1820) (7c5614a)
• Fix response type of ollama (#1850) (41e8cae)
• Handle missing runs to avoid out of range exception (#1844) (4002de1)

Feature

• Support audio input (#1763) (1557e7c)
• markdown: Add formatting & improve inline support (#1804) (861abcd)
• Maximum image size for Vlm models (#1802) (215b540)

Fix

• docx: Ensure list items have a list parent (#1827) (d26dac6)
• msword_backend: Identify text in the same line after an image #1425 (#1610) (1350a8d)
• Ensure uninitialized pages are removed before assembling document (#1812) (dd7f64f)
• Formula conversion with page_range param set (#1791) (dbab30e)

Documentation

• Update readme and add ASR example (#1836) (f3ae302)
• Support running examples from root or subfolder (#1816) (64ac043)

Feature

• Make Page.parsed_page the only source of truth for text cells, add OCR cells to it (#1745) (7d3302c)
• Support xlsm files (#1520) (df14022)

Fix

• Pptx line break and space handling (#1664) (f28d23c)
• asciidoc: Set default size when missing in image directive (#1769) (b886e4d)
• Handle NoneType error in MsPowerpointDocumentBackend (#1747) (7a275c7)
• Prov for merged-elems (#1728) (6613b9e)
• tesseract: Initialize df_osd to avoid uninitialized variable error (#1718) (e979750)
• Allow custom torch_dtype in vlm models (#1735) (f7f3113)
• Improve extraction from textboxes in Word docs (#1701) (9dbcb3d)
• Add WEBP to the list of image file extensions (#1711) (a2b83fe)

Documentation

• Update vlm models api examples with LM Studio (#1759) (0432a31)
• Add open webui (#1734) (49b10e7)

Fix

• Remove typer and click constraints (#1707) (8846f1a)

Documentation

• Flash-attn usage and install (#1706) (be42b03)

Feature

• Simplify dependencies, switch to uv (#1700) (cdd4018)
• New vlm-models support (#1570) (cfdf4ce)

Feature

• Add visualization of bbox on page with html export. (#1663) (b356b33)

Fix

• Guess HTML content starting with script tag (#1673) (984cb13)
• UnicodeDecodeError: 'utf-8' codec can't decode byte 0xd0 in position 0: invalid continuation byte (#1665) (51d3450)

Documentation

• Fix typo in index.md (#1676) (11ca4f7)

Feature

• ocr: Auto-detect rotated pages in Tesseract (#1167) (45265bf)
• Establish confidence estimation for document and pages (#1313) (9087524)

Fix

• Fix ZeroDivisionError for cell_bbox.area() (#1636) (c2f595d)
• integration: Update the Apify Actor integration (#1619) (14d4f5b)

Feature

• Add textbox content extraction in msword_backend (#1538) (12a0e64)

Fix

• Fix issue with detecting docx files, and files with upper case extensions (#1609) (f4d9d41)
• Load_from_doctags static usage (#1617) (0e00a26)
• Incorrect force_backend_text behaviour for VLM DocTag pipelines (#1371) (f2e9c07)
• pypdfium: Resolve overlapping text when merging bounding boxes (#1549) (98b5eeb)

Feature

• Improve parallelization for remote services API calls (#1548) (3a04f2a)
• Support image/webp file type (#1415) (12dab0a)

Fix

• ocr: Orig field in TesseractOcrCliModel as str (#1553) (9f8b479)
• settings: Fix nested settings load via environment variables (#1551) (2efb7a7)

Documentation

• Add advanced chunking & serialization example (#1589) (9f28abf)

Fix

• AsciiDoc header identification (#1562) (#1563) (4046d0b)
• Restrict click version and update lock file (#1582) (8baa85a)

Fix

• Add smoldocling in download utils (#1577) (127e386)
• HTML: Handle row spans in header rows (#1536) (776e7ec)
• Mime error in document streams (#1523) (f1658ed)
• Usage of hashlib for FIPS (#1512) (7c70573)
• Guard against attribute errors in TesseractOcrModel del (#1494) (4ab7e9d)
• Enable cuda_use_flash_attention2 for PictureDescriptionVlmModel (#1496) (cc45396)
• Updated the time-recorder label for reading order (#1490) (976e92e)
• Incorrect scaling of TableModel bboxes when do_cell_matching is False (#1459) (94d66a0)

Documentation

• Update links in data_prep_kit (#1559) (844babb)
• Add serialization docs, update chunking docs (#1556) (3220a59)
• Update supported formats guide (#1463) (3afbe6c)