Quasar System for RAG-based Question Answering

This technical report presents the Quasar system, designed for question answering (QA) over heterogeneous data sources, including unstructured text, structured tables, and knowledge graphs. The system utilizes a RAG-based architecture that integrates evidence retrieval and answer generation through a moderate-sized language model. Unique to Quasar is its capability for question understanding, which refines user input for improved evidence retrieval, and a re-ranking mechanism that filters retrieved evidence to enhance answer quality. The report details the methodology, which consists of four main stages: Question Understanding (QU), Evidence Retrieval (ER), Re-Ranking & Filtering (RF), and Answer Generation (AG). Experiments conducted across three benchmarks demonstrate that Quasar achieves high answering quality comparable to large GPT models while significantly reducing computational costs and energy consumption. The findings emphasize the importance of structured intent representation and iterative re-ranking in optimizing performance for diverse question types.