Structured RAG for Answering Aggregative Questions

The introduction of S-RAG marks a significant advancement in the field of question answering, particularly for aggregative queries that necessitate comprehensive information gathering from extensive document sets. Traditional RAG systems have struggled with such queries, often focusing on smaller, more specific segments of data. By creating a structured representation of the corpus at ingestion and translating queries into formal structures at inference, S-RAG enhances the ability to reason over large datasets. The validation of this approach is supported by the introduction of two new datasets, HOTELS and WORLD CUP, specifically designed for aggregative queries. Experiments have demonstrated that S-RAG outperforms both conventional RAG systems and long-context LLMs, indicating its potential to transform how complex questions are answered in AI applications.