Docling’s latest release adds an optional way to retrieve images embedded in PDF files, alongside fixes that preserve pictures, captions and scholarly tables during conversion. The document-processing tool also gains speech-recognition support through mlx-audio and a MiniCPM-OCR preset. The image extraction setting must be enabled by the user; the update does not switch it on automatically.
Docling, a tool that turns documents into forms usable in AI workflows, added extraction of embedded PDF pictures in release 2.149.0 on 8 November. Users choose whether to activate the capability. Image handling also changes for pictures in modes that cannot be saved as PNG: those pictures are now retained. These changes concern obtaining and keeping visual content during document conversion.[1]
The update keeps captions attached to their pictures when converting web documents. For scholarly publications in the JATS format, it preserves tables supplied as images and corrects handling of their footnotes. A separate adjustment to reading order prevents horizontal expansion into another document element. Together these are changes to how existing page components survive conversion, including the links between a picture, its accompanying text and surrounding elements.[1]
Speech-to-text processing gains an additional option through mlx-audio support. MiniCPM-OCR is added as a preset in the configuration for models that process both pictures and language. The command-line interface has also been reorganised. Text conversion receives targeted repairs: LaTeX numbered lists retain item numbering, while the CSV reader follows a delimiter declared by Excel on the first line. The release combines these processing changes with its optional image-extraction feature.[1]
