RAR-Agent: 법률 질의응답을 위한 근거 보강형 검색 시스템

Hallucination and outdated knowledge in large language models critically undermine their reliability and applicability in specialized domains such as law and medicine, where factual accuracy is essential. While Retrieval-Augmented Generation (RAG) has been proposed as a mitigation strategy, its effectiveness in the legal domain is often hindered by lexical mismatches, which impede the accurate retrieval of highly relevant external knowledge. Although several studies have explored query formulation–based approaches to address this issue, additional training costs and hallucination during the retrieval phase remain persistent challenges. In this paper, we propose RAR-Agent (Rationale-Augmented Retrieval Agent) to overcome these limitations. RAR-Agent employs a Chain-of-Thought and Rationale-based query formulation technique, combined with a Reciprocal Rank Fusion and Reranker-based filtering mechanism, to alleviate lexical mismatch problems and effectively suppress hallucination during retrieval. Furthermore, to precisely evaluate the agent’s factual accuracy, we constructed the KL-BQA (Korean Legal Binary Question-Answering) benchmark. The proposed model achieved superior performance on both the KL-BQA and KL-RQA benchmarks.