Beyond Text: Building Multimodal RAG Agents That Read, Reason, and Retrieve

Most production RAG systems remain stubbornly text-centric - yet the documents, dashboards, and data stores that enterprises rely on are richly multimodal. A single PDF might interleave dense paragraphs with scanned tables, architectural diagrams, and annotated screenshots. This talk bridges that gap with a practical architecture for Multimodal RAG Agents: systems that retrieve and reason over heterogeneous content spanning text, images, charts, and structured data. We walk through the full pipeline - from multimodal embedding strategies and cross-modal indexing to agentic orchestration patterns that dynamically route queries to the right retrieval substrate. Along the way, we tackle the real engineering challenges: preserving layout context when chunking visually rich documents, fusing retrieved multimodal evidence before generation, and evaluating agent performance when ground truth spans multiple modalities. Attendees will leave with a clear mental model for designing multimodal retrieval pipelines, deciding when to use agents, and building systems that move beyond chat over documents toward AI assistants that can understand and operate across real enterprise knowledge.
Format: TECHNICAL TALK
Location: ROOM A
Speakers: Catalin Hanga
Level: Intermediate
Audience: Registered participants