A new AI pipeline called RegulaRAG is designed to stop large language models from hallucinating their way through dense safety regulations.
Researchers built RegulaRAG as a retrieval-augmented generation (RAG) system, a setup where an AI model pulls relevant text from a document before answering instead of relying on memory alone. It pairs a step called SmartChunking, which enriches paragraphs and tables with their cross-references using graph traversal, with a second step called Smart Retrieve and Rerank. The team tested it against five baseline RAG systems on a hand-curated dataset covering every scenario in UN Regulation No. 152, the standard governing automatic emergency braking systems (AEBS). On a custom scoring metric, RegulaRAG hit an average of 82.99, beating the next-best system by 43 percent, while using 14,000 to 25,000 tokens per query versus up to 500,000 for graph-heavy rivals.
This matters because automakers increasingly lean on AI to generate the test scenarios that prove a vehicle meets safety rules, and regulations are long, cross-referenced, and unforgiving of sloppy citation. RegulaRAG also held its accuracy steady as researchers added distractor content to simulate a growing body of regulations, while the competing systems degraded sharply.
Still, this is one case study on one regulation. Whether the approach holds up across the sprawling, inconsistent style of global auto safety codes is the real test, and nobody has run that yet.