LLM-Driven Retrieval, Debate, and Verification for Robust Table‐to‐Knowledge‐Graph Matching

Tabular data is one of the most common data sources on the internet and is widely used in various data analytics tasks. Identifying semantic concepts within tables is often a critical component of these pipelines, yet it remains a challenging task to automate. To address this problem, we present RAGDify, a large language model (LLM)-based system designed for the Cell Entity Annotation (CEA) task. Our system employs a three-step pipeline inspired by Retrieval-Augmented Generation (RAG) and advanced reasoning techniques: (1) retrieving context-aware candidate entities, (2) engaging in a debate-like evaluation to compare top candidates, and (3) applying chain-of-verification-inspired prompting to validate the final entity match. We propose RAGDify as a solution for the SemTab’25 challenge, targeting the key challenges inherent in automating the CEA task.