A new retrieval method wants AI systems to stop losing track of what's actually in your documents.
Researchers built a system called Hyper-M2RAG that changes how retrieval-augmented generation, the technique that lets an AI model pull facts from an external document before answering, connects the dots in mixed-media files. Most current systems model a document as a simple graph, linking content in pairs, which breaks down when a chart, its caption, and an underlying data table all describe the same thing at once. Hyper-M2RAG instead builds a hypergraph, where a single hyperedge can tie together text, images, and tables that belong together, and uses anchor nodes at page boundaries to patch context locally instead of re-scanning entire documents. The team reports it beats existing methods on multimodal benchmarks for both retrieval accuracy and generation coherence, and has published the code on GitHub.
Real documents are not sets of isolated facts. A chart, its footnote, and a number buried three paragraphs later are one idea split across formats, and treating them as separate pairwise links loses that connection. If hypergraph modeling holds up outside the researchers' own benchmarks, it could meaningfully improve AI search over financial reports, technical manuals, and other documents built from mixed text and images.
The gains here are self-reported against benchmarks the team chose, and hypergraphs plus anchor-based refinement add real engineering complexity for anyone trying to ship this outside a research paper.