Abstract

Weaver is a modular pipeline that dynamically combines SQL and Large Language Models (LLMs) for advanced table-based question answering. Unlike rigid approaches, Weaver generates flexible execution plans that use SQL for structured data operations and LLMs for semantic reasoning, automatically deciding the best tool for each subtask. Our method consistently outperforms state-of-the-art approaches across four major TableQA datasets while reducing API costs and improving accuracy through intelligent query decomposition.