GINGER: Grounded Information Nugget-Based Generation of Responses
Topics: AI (Deep Learning), Krisztian Balog, LLMO / GEO, Retrieval Augmented Generation (RAG)
This paper introduces GINGER (Grounded Information Nugget-Based Generation of Responses), a novel approach for generating accurate and verifiable responses in retrieval-augmented generation systems. The system works by breaking down retrieved passages into atomic information units called “nuggets,” which are then clustered, ranked, and synthesized into coherent responses. The approach demonstrated superior performance in the TREC RAG’24 evaluation, particularly in ensuring factual accuracy and source attribution while maintaining response fluency.