2026-07-15 –, Memorial Hall
Scientific breakthroughs don’t happen in plain text, they live inside multi-column research papers, dense data tables, and intricate simulation diagrams. Yet the moment standard AI and Retrieval-Augmented Generation (RAG) pipelines encounter these layouts, they fail. Tables are flattened into meaningless strings. Figures are ignored. The structural signals that drive scientific reasoning disappear.
In this talk, we show how to rescue scientific knowledge from the “text-flattening” trap using Docling, an open-source document understanding library designed to preserve layout, hierarchy, and element boundaries. Instead of reducing everything to text, we treat tables, figures, and sections as first-class data structures. Attendees will experience a live demo of a realistic scientific R&D workflow: uploading multiple dense technical PDFs, executing cross-document natural language queries, and successfully retrieving synthesized insights from text, structured tables, and scientific images
When analyzing simulation reports, experimental summaries, and technical journals, researchers must extract evidence, compare results, and validate claims across multiple sources simultaneously. To enable this rigorous level of analysis, we will walk through the technical implementation of a structure-aware ingestion pipeline.
Using Python, we will demonstrate an architecture that decomposes document layouts into distinct semantic elements - sections, tables, and figures. This approach preserves experimental results as queryable data structures and converts diagrams into searchable semantic signals, all while maintaining the strict document hierarchy required for context-aware retrieval. Building on this foundation, we detail the construction of a hybrid retrieval system that actively supports:
- Cross-document comparison
- Numeric reasoning over extracted tables
- Linking textual claims to supporting figures
- Combining text, structured data, and visual insights in a single grounded response
Outline
- The Scientific Workflow Challenge
- Structure-Aware Ingestion
- Preserving and Querying Tables
- Visual Representation and Linking
- Live Demo & Multimodal Retrieval
- Q&A
Participants will gain a practical design pattern for building multimodal, structure-preserving retrieval systems that strengthen scientific reasoning and data-driven analysis.
Mahima Arora is a Senior Data Scientist on the Data & AI team at Red Hat, specializing in Generative AI applications. She develops AI-powered solutions that enhance efficiency and effectiveness, leading initiatives to optimize AI systems for greater impact. Passionate about open source, Mahima actively explores emerging tools and technologies to drive innovation and knowledge sharing, and has presented her work at PyData Amsterdam 2025 and PyCon India 2025.
Aarti Jha is a Principal Data Scientist at Red Hat, where she leads the development of AI-driven solutions that streamline internal operations and reduce costs. She has more than seven years of experience designing and deploying machine learning and generative AI solutions across multiple industries. A frequent speaker at developer and data science conferences, she shares practical insights on applied AI, LLMs, and building AI systems that deliver measurable business value.