Document Retrieval Beyond Text Embeddings

This thesis explores modern document retrieval methods that go beyond standard text embeddings. It argues that retrieval systems should make better use of readily available but often overlooked information, notably by exploiting new forms of associative links between query terms and document content. Part I introduces Visual Document Retrieval (VDR), in which document pages are indexed directly from their images, allowing text, layout, tables, figures, and other visual cues to contribute jointly to document representations. Part II investigates how contextualizing document passages and search queries through structural links or LLM parametric knowledge can improve representation construction. Together, these contributions demonstrate that moving beyond isolated text embeddings can make retrieval systems more effective, efficient, and better suited to real-world information needs.