FT-RAG Framework Enhances Table Reasoning in LLMs

The newly introduced FT-RAG framework enhances Retrieval-Augmented Generation (RAG) specifically for intricate table reasoning. Traditional RAG systems often falter with structured tabular information due to their broad retrieval granularity and lack of deep semantic comprehension. FT-RAG overcomes these challenges by breaking down tables into entry-level semantic components to form a structured graph. It utilizes a structural neighbor expansion approach to identify semantically related entities during graph retrieval, followed by multi-modal fusion for context integration. To tackle the issue of limited datasets, the authors present Multi-Table-RAG-Lib, a benchmark featuring 9,870 complex QA pairs that necessitate the integration of multiple tables and text-table fusion. The paper can be accessed on arXiv with ID 2605.01495.