RAG-Anything: All-in-One RAG Framework (arxiv.org)

RAG-Anything is an open-source, all-in-one Retrieval-Augmented Generation framework that extends RAG beyond text-only corpora to handle multimodal knowledge sources (text, images, tables, math, etc.). The paper reframes multimodal documents as interconnected knowledge entities and introduces a unified representation that closes the gap between existing RAG systems and real-world, heterogeneous information environments. The authors report substantial improvements on challenging multimodal benchmarks, with especially large gains on long documents where conventional, text-centric RAG methods break down. Technically, RAG-Anything builds a dual-graph representation to capture both cross-modal relationships and intra-textual semantics, and pairs that with a cross-modal hybrid retrieval strategy that blends structural knowledge navigation (graph-based traversal of entities/relations) with semantic matching (embedding-based retrieval). This lets the system retrieve and reason over evidence that spans modalities—e.g., linking a table cell to a figure and a paragraph—without stitching together separate pipelines. The result is a single, coherent architecture that reduces fragmentation, improves retrieval precision on heterogeneous content, and enables more faithful multimodal reasoning for downstream LLM generation. The framework and code have been open-sourced for community use and extension.