SCMRAG 2.0: Efficient and Scalable Multi-hop Graph RAG with Multimodal Knowledge-Graphs and Agentic Self-Correction

Rishabh Agrawal, Uday Devulapalli, Apurva Narayan

Keywords: retrieval-augmented generation, multimodal knowledge graphs, graph retrieval, agentic self-correction, multihop reasoning

TL;DR: SCMRAG 2.0 is a multimodal graph RAG framework that fuses language- and embedding-linked knowledge graphs with optimized retrieval and agentic self-correction to yield faster, more accurate, and reliable multimodal reasoning and generation.

Abstract: We present SCMRAG 2.0, a next-generation retrieval-augmented generation framework that extends SCMRAG's self-corrective design to truly multimodal knowledge graph and agentic reasoning. SCMRAG 2.0 introduces four advances: (1) a Multimodal Knowledge Graph (MMKG) whose nodes unify text, image, audio, and structured artifacts into claim-anchored representations; (2) dual linkage among nodes via both symbolic/language relations and learned cross-modal embedding edges to capture latent correspondences; (3) an optimized graph-embedding and -retrieval algorithm that jointly prunes and composes across modalities, improving recall–precision trade-offs and significantly improving latency while preserving multihop coverage; and (4) a multimodal agentic self-correction loop that audits evidence support, issues modality-aware follow-ups, and iteratively amends the MMKG and context with newly found cross-modal evidence. By aligning language-level structure with vector-space signals and enabling agentic critique and repair, SCMRAG 2.0 mitigates outdated context, incomplete reasoning chains, and hallucinations common in text-only graph RAG. Experiments on multihop QA and creative-assistance tasks mixing language and vision show consistent gains in retrieval precision, factuality, and compute efficiency over SCMRAG and strong RAG baselines, indicating a scalable path to reliable multimodal RAG systems for knowledge-intensive and creative workflows.

Area: Generative and Agentic AI (GAAI)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1018