AI
GEMS-RAG: Graph-grounded Engineering Multimodal Standards Retrieval-Augmented Generation
Publication: cs.AI updates on arXiv.org
Author: Abdul Hannan Azad, Irvin Wang, Jiwan Jiang, Chenchen Kuai, Bofeng Cao, Keshu Wu, Yuzhou Wang, Xiaopeng Li, Yang Zhou
Source published: 27/10/2026, 04:00:00 UTC
Briefing created: 22/11/2026, 16:38:04 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 28/10/2026, 02:33:25 UTC.
Briefings
A new framework called GEMS-RAG combines hierarchical parsing, fine-grained visual entity extraction, and graph-grounded retrieval to improve engineering standards analysis. Evaluated on the MUTCD, it outperforms text-only, multimodal, and graph-based baselines in question-answering performance. The system decomposes complex figures into linked entities and uses graph expansion to find relevant evidence across text, visuals, and document structure.
Key points
- GEMS-RAG improves retrieval effectiveness and question-answering performance on the MUTCD compared to text-only, multimodal, and graph-based RAG baselines.
- The framework decomposes complex figures into meaningful visual entities linked to textual provisions, engineering concepts, and regulatory rules.
Context
Engineering standards are challenging knowledge sources for LLMs due to their hierarchical, multimodal, and densely connected nature. Existing RAG methods struggle with joint reasoning over text, visuals, and document structure. This finding matters as it offers a general framework for reasoning over engineering codes in safety-critical applications.
Limitations
Evaluation is limited to the MUTCD; generalizability to other engineering standards remains untested. Results are from a single preprint snapshot with no validation on real-world deployment or long-term performance.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/11/2026, 16:38:04 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeidqb5zwwd5ntktmvwcn2mza4xvhezn24iyylycx4dsyvchmtf2j2a