Abstract

Retrieval-augmented generation (RAG) faces challenges related to factualcorrectness, source attribution, and response completeness. To address them, wepropose a modular pipeline for grounded response generation that operates oninformation nuggets-minimal, atomic units of relevant information extractedfrom retrieved documents. The multistage pipeline encompasses nugget detection,clustering, ranking, top cluster summarization, and fluency enhancement. Itguarantees grounding in specific facts, facilitates source attribution, andensures maximum information inclusion within length constraints. Extensiveexperiments on the TREC RAG'24 dataset evaluated with the AutoNuggetizerframework demonstrate that GINGER achieves state-of-the-art performance on thisbenchmark.