Abstract
Public servants and auditors must often identify precise legal evidence across large collections of laws, decisions, audit reports, and technical documents. However, institutional search systems typically rely on rigid metadata or exact keyword matching, making it difficult to answer natural-language questions whose evidence is distributed across heterogeneous sources. This article presents a Retrieval-Augmented Generation (RAG) framework for legal corpora that combines hybrid retrieval, keyword extraction, metadata-driven filtering, and multi-hop retrieval for legal-administrative question answering. The study follows a staged design. Stage I establishes a keyword-augmented hybrid RAG baseline that integrates dense and lexical retrieval with automatic keyword extraction. Stage II extends this baseline with query decomposition, LLM-guided metadata extraction for dynamic filtering, and domain-specific cross-encoder reranking. We evaluate the framework on two corpora: a test set with more than 7,000 legislative and administrative documents and a validation corpus with more than 333,000 PDFs and 10 terabytes of data across five institutional knowledge bases: legislation, decisions, audit reports, prosecutor opinions, and portal files. Results show that keyword-guided hybrid retrieval improves evidence coverage and ranking quality, while structured multihop control further improves retrieval coverage and answer correctness in heterogeneous, high-stakes legal corpora.