LF AI & Data Foundation

Please join us in congratulating Peter W. J. Staar on his election as the new Chair of the LF AI & Data Foundation Technical Advisory Council (TAC)! In his first blog as TAC Chair, Peter shares his vision for the future of open source AI infrastructure, highlighting why the industry's next challenge is building an open, interoperable context layer that enables AI systems to better understand and reason over real-world knowledge. Read the blog: https://lnkd.in/gcwJkFJ8 Congratulations, Peter! We're excited to see the community continue advancing open collaboration together. Join the LF AI & Data Foundation: https://lnkd.in/g589d3DU #OpenSource #AI #ArtificialIntelligence #LFAIData #OpenAI #MachineLearning

"The world's knowledge lives in formats — PDF, HTML, Word, LaTeX, MP3, MP4 — that were designed for rendering or playback, not for understanding." (from the DocLang GitHub readme) Meet DocLang, an emerging open standard for AI-native documents 💡 LF AI & Data Foundation announced the DocLang Specification Working Group on June 9, 2026. It operates under the Joint Development Foundation’s open governance model, with contributors including IBM, NVIDIA, Red Hat, ABBYY, HumanSignal, and Forgis. What interests me is not only the format itself, but the pre-standardization work happening in public: shared development, industry feedback, and early quality guidance for how agents consume information. DocLang uses a constrained XML-based vocabulary designed for token efficiency. Its direction is to stay below roughly 1,000 syntax tokens, avoid attribute-heavy markup where possible, and use a representation that maps more naturally to LLM tokenizers. Elements can carry semantic tags, page coordinates, bounding boxes, and reading order. Tables keep their grid structure through 𝗢𝗧𝗦𝗟 (optimised table-structure language). Headings carry levels. Content remains traceable back to where it appeared in the source. All in support of keeping a structured overview of consumed content. Benefits I see: ▪️ Data access quality determines output quality and efficient consumption ▪️ Humans need organized information access just as much ▪️ Metadata documentation as a bridge towards compliance through traceability and transparency contributions DocLang is weeks old, and standards live or die on who implements them. The vision is a broadly adopted international standard for AI-ready documents, providing a consistent representation for both humans and machines. How I’d use it practically: Docling, the IBM open-source document processing toolkit, already supports DocLang output. For my own workflows, that means I could parse PDFs into structured, machine-readable data for my AI workflows and knowledge base instead of relying on flat extracted text. (Audio and video files are already supported in Docling — has anybody tried how well this works so far? Especially interested in accessing video content.) DocLang feels like one piece of a larger shift that's been going on for a little while now. 𝗹𝗹𝗺𝘀.𝘁𝘅𝘁 gives agents a map of a website. 𝗠𝗖𝗣 standardizes how they reach tools and data. 𝗖𝟮𝗣𝗔 attaches provenance signals to digital content. Different directions, same underlying recognition: 𝗛𝘂𝗺𝗮𝗻-𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗰𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝗼𝗻 𝗻𝗲𝗲𝗱𝘀 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁𝗹𝘆 𝗱𝗲𝘀𝗶𝗴𝗻𝗲𝗱 𝗶𝗻𝘁𝗲𝗿𝗳𝗮𝗰𝗲𝘀, 𝗯𝗼𝘁𝗵 𝘄𝗮𝘆𝘀. ➡️ How do we best design one world of knowledge that both humans and machines can read? Would love to hear your thoughts. Find links in the comments. Collab with Claude & Codex.

Your RAG system is only as good as the documents it can actually understand. Before adopting Docling, AskIBM could search HTML pages, but not the valuable information hidden inside attached PDFs, DOCX files, presentations, and images. After integrating Docling: ✅ 250,000 new searchable passages added (10% of the entire knowledge base) ✅ 5–6% of employee queries now return document-based results that previously couldn't be found ✅ ~60,000 retrievals every day across internal API consumers Sometimes improving AI isn't about changing the model. It's about giving it access to the knowledge that already exists. Read how IBM's CIO team did it with LF AI & Data project: https://lnkd.in/grJQ6xXZ #EnterpriseAI #RAG #OpenSource #Docling #LFAIData

Congratulations to Peter W. J. Staar on being elected Chair of the LF AI & Data Foundation Technical Advisory Council! We're excited for what's ahead and can't wait to see Peter help grow our amazing open source AI community. Looking forward to working together to advance open collaboration and innovation in AI. Congratulations, Peter!

Open source innovation happens because of its community. A huge thank you to the 27 contributors who helped deliver ONNX v1.22.0, including 16 first-time contributors. Together, the community introduced new generative AI capabilities, strengthened supply-chain security, improved developer tooling, and continued advancing one of the industry's leading standards for machine learning interoperability. Read about the release and discover how to get involved, written by Andreas Fehlner: https://lnkd.in/eyREY6Vx #OpenSource #ONNX #Community #AI #MachineLearning #LFAIData

This is a fantastic spotlight on how Docling bridges the gap between raw, unstructured enterprise data and powerful AI workflows. Cedric Clyburn does an excellent job demonstrating the practical versatility of the tool across several key use cases: Complex Document Parsing: He highlights how to handle challenging inputs like multi-page tables, scanned PDFs, and image-heavy layouts, showing how the library preserves structural integrity where simple parsers often fail. Efficiency at Scale: Cedric covers the cost and performance benefits of running Docling locally, proving that high-quality document conversion doesn't require expensive proprietary services or heavy GPU reliance. Agentic & Chunkless RAG: He demonstrates advanced patterns, including chunkless RAG—using document outlines for retrieval—and integrating Docling with the Model Context Protocol (MCP) to empower AI agents to interact with document structures directly. It is great to see the community utilizing these features to build more reliable, context-aware AI systems. Whether you are building RAG pipelines or preparing datasets for fine-tuning, this session offers a comprehensive look at how open-source tools can solve real-world document engineering hurdles. 📺 https://lnkd.in/eZD_RfGz #Docling #RAG #GenerativeAI #DocumentEngineering #DocumentAI #OpenSource #AIApplication #ModelContextProtocol Red Hat IBM LF AI & Data Foundation

Most enterprise knowledge isn't missing. It's trapped inside PDFs, PowerPoint decks, Word documents, and images. IBM's CIO team integrated Docling into AskIBM and the results speak for themselves: 📄 250,000 previously invisible document passages unlocked 👥 Searchable across 280,000 IBM employees 📈 ~4,000 document-based results delivered every day By making attachments AI-ready, IBM dramatically expanded the knowledge available to its internal AI assistant, without replacing its existing architecture. This is what enterprise AI looks like when document understanding becomes a first-class capability. Read the case study below 👇 👇 Special thanks to Peter W. J. Staar, Carol Chen, Mingxuan Z. and more IBM community members for supporting this case study! #Docling #OpenSourceAI #EnterpriseAI #RAG #LFAIData #IBM

ONNX v1.22.0 is here! Authored by Andreas Fehlner, this latest release introduces major improvements that make ONNX even better for modern AI development, including: ✅ Native attention operators for generative AI and LLMs ✅ WebAssembly support for in-browser model validation ✅ Supply-chain security with SLSA Level 2 provenance and embedded SBOMs ✅ Modernized, more reliable builds across platforms Huge thanks to the 27 contributors, including 16 first-time contributors, who helped make this release possible. Read what's new: https://lnkd.in/eyREY6Vx #ONNX #OpenSourceAI #GenerativeAI #MachineLearning #LLM #LFAIData

🚀 Docling meetst IBM Bob: From Unstructured Documents to Structured Data 🚀 What if you could go from a complex document to a working AI-powered application in a fraction of the time? That's exactly what you can do by combining IBM Bob with Docling. Every enterprise has documents—purchase orders, invoices, contracts, reports—but turning them into structured, usable data is often one of the biggest bottlenecks in building AI solutions. This is where the combination really shines: ✨ Docling intelligently extracts structured data from complex documents. ✨ IBM Bob helps accelerate application development with AI-assisted, spec-driven development. The result? Less time wrestling with document parsing and boilerplate code, and more time building solutions that deliver real business value. The latest tutorial (https://lnkd.in/eD_CnzYM) demonstrates this beautifully by walking through the creation of a purchase order processing application—from document ingestion to structured data extraction and application development. 📖 Explore it yourself: 👉 Docling: https://lnkd.in/d4UT-6_2 👉 Docling-SaaS: https://lnkd.in/d__EknDi 👉 Docling + IBM Bob: https://lnkd.in/eD_CnzYM #IBM #IBMBob #Docling #AI #GenerativeAI #DocumentAI #Developer #EnterpriseAI #Automation #RAG #AgenticAI #LLM