Abstract
Legal practitioners increasingly rely on automated document summarization to navigate complex contracts, case laws, and statutory texts. Despite significant progress in large language models (LLMs), hallucinations—fabricated facts, misinterpretations, or inaccurate citations—remain a major barrier to adoption in high-stakes legal contexts. Retrieval-Augmented Generation (RAG) has emerged as a promising approach to mitigate hallucinations by grounding model outputs in verified documents. This review evaluates the landscape of RAG architectures for legal document summarization, highlighting their design, efficacy, benchmarking methods, and limitations. We synthesize evidence showing that while RAG reduces hallucination rates by 20–60%, its performance is highly sensitive to corpus indexing, retrieval quality, prompt design, and domain specificity. We outline future research opportunities spanning multi-modal retrieval, chain-of-verification mechanisms, and constitutional-style safety constraints for legal AI.
