UniversalRAG: Retrieval-Augmented Generation over Corpora of Diverse Modalities and Granularities

In three linesUniversalRAG is a multi-modal RAG framework that retrieves and integrates knowledge from heterogeneous sources (text, images, videos) at variable granularities. It introduces modality-aware routing to avoid intra-modal bias and organizes each modality into granularity levels. Validated on 10 benchmarks, it outperforms single-modality and unified baselines.

Summary generated by Claude — human-verified