Using Retrieval-Augmented Generation to Improve Access to Legal Documents

Access to legal and regulatory documents is critical for efficient government operations, yet current search systems rely on rigid metadata and lack contextual understanding. This paper presents a Retrieval-Augmented Generation (RAG) frame-work that integrates hybrid retrieval with a keyword extraction module to improve precision by identifying key terms from the query. The system ingests and indexes over 7,000 Brazilian legislative documents, enabling natural-language question answering through multiple large language models (LLMs). We evaluate nine RAG configurations combining open-source and proprietary models using recall, MRR, BERTScore, and METEOR. The best setup in the retrieval step (Mistral + bge-m3 embeddings) achieved 0.852 recall, while Llama 4 Maverick obtained 0.785 BERTScore and 0.48 METEOR. These results demonstrate that RAG architectures can substantially improve factual accuracy and retrieval coverage in legal-document access. Among the open-source models showed particularly promising performance, highlighting the growing competitiveness of community-driven LLMs in specialized legal retrieval tasks.