PRISM-RAG: Multimodal Hypergraph Retrieval-Augmented Generation for Tobacco Product and Legislative Policy Reasoning

The disambiguation of semantically similar statutory text across jurisdictions is a retrieval problem that existing methods do not solve. This inter-context conflict can steer generative models toward confidently produced answers grounded in topically relevant but jurisdictionally incorrect, sources. Tobacco and nicotine control regulations vary across the United States, from federal, state, and municipal statutes and administrative codes that may share similar language. Thus, robust reasoning requires identifying which jurisdiction's law governs a given product, not merely retrieving semantically similar or relevant text for a given query. This is critical because emerging nicotine products (e.g., e-cigarettes and oral nicotine products) may circumvent existing regulations due to product definitions, and governments, agencies, and public health researchers and practitioners struggle to know what products are subject to which regulation.

State-of-the-art (SOTA) document retrieval-augmented generation (RAG) methods break down in this task, rife with inter-context conflict, for two reasons. First, they provide no mechanism to ground a query in a product image, and second, embedding-similarity or entity-based retrieval cannot distinguish a jurisdiction's statute from another jurisdiction's similar text. To expose this failure case and provide a benchmark for evaluating solutions to it, we introduce the Nicotine Product and Regulation Image-and-Text Surveillance Multimodal (NicoPRISM) dataset, developed by our team of computer scientists and public health researchers. NicoPRISM comprises 161,563 images from web and social media sources paired with seven structured captions, a curated knowledge base of product, health, and static legislative documents spanning 13 US exemplar jurisdictions, and 1,495 validated question-answer pairs organized into two benchmark tasks: policy compliance QA and product knowledge QA.

To address the jurisdictional disambiguation problem directly, we propose PRISM-RAG, a multimodal hypergraph RAG framework that constructs a compact hypergraph over images, captions, and entities from heterogeneous document data without any large language model calls at index time. PRISM-RAG grounds every query in a product image, then routes retrieval through a jurisdiction-aware context assembly mechanism that guarantees statutory text from the queried jurisdiction reaches the language model by construction, independent of embedding-space topology. PRISM-RAG retrieves passages from the correct jurisdiction in 93.9% of policy compliance queries, a 48.6 percentage point advantage over standard RAG (p < 0.001), while requiring zero large language model calls at index time and a single call at query time. Across metrics measuring ground-truth keyword retrieval, semantic similarity, jurisdiction retrieval accuracy, compliance label accuracy, and context coverage, several of which expose limitations in SOTA document RAG methods, PRISM-RAG is competitive with and outperforms current SOTA document RAG frameworks while minimizing LLM calls at both index and query time.