DualRAG Architecture: Structured Retrieval with Layered Knowledge Models

Michal Podpora, Marek Baranowski, Aleksandra Kawala-Sterniuk, Mariusz Pelc

Abstract: DualRAG is a novel architecture for Retrieval-Augmented Generation (RAG) that introduces a two-layer knowledge representation to improve relevance assessment and context construction. Each knowledge chunk is divided into two parts, separated by a delimiter ("+="): a concise, relevance-focused descriptor and a detailed, content-rich segment. During retrieval, only the relevance descriptor is used for scoring and ranking candidate chunks. After selection, both parts are combined to construct a prompt with greater contextual depth and precision. This decoupling of retrieval scoring from generative input enables more accurate filtering without compromising the expressiveness of retrieved content. The approach is particularly suited to domains requiring fine-grained distinction and layered factual detail, such as environmental regulations or technical documentation. DualRAG improves retrieval quality while maintaining compatibility with existing RAG pipelines and model infrastructures.

External IDs:dblp:conf/icaart/PodporaBKP26