Deep Dive into Visual Grounding: Measuring Bounding Boxes and Citation Accuracy in Document AI

ExtractBench is an open benchmark for schema-guided document extraction. In this video, we do a deep dive into visual grounding and measuring word- and page-level spatial citations. When AI agents execute autonomous workflows, auditability is critical. An extraction system cannot just return a text value; it must prove exactly where on the physical page that value originated. We score spatial grounding deterministically using word- and page-level bounding box overlap. This tests whether models can emit verifiable visual citations back to the source document for enterprise compliance and huma