Enhancing Legal Document Building with Retrieval-Augmented Generation

Legal document building refers to the process of producing a legal textual document following a predefined schema with the support of digital, automated tools. Such systems must balance two fundamental requirements: providing targeted drafting assistance while preserving judicial autonomy and decision-making authority, and systematically leveraging existing legal document corpora to enhance consistency and quality in legal documentation. In this paper, we propose a document builder architecture, called JusBuild, designed to assist and support legal practitioners in drafting new legal documents. JusBuild supports the document assembly process by relying on a predefined legal document template and on a corpus of past legal documents. The key features of JusBuild are: (i) the use of a Conditional Random Field (CRF) model for the supervised segmentation of legal documents into functional sections according to a document template; (ii) a vector database storing segmented sections and their semantically meaningful vector representations for efficiently performing semantic search for suggestions retrieval; (iii) the suggestion, at drafting time, of relevant precedent sections retrieved from the vector database and of new, AI-generated sections, using a Large Language Model and Retrieval-Augmented Generation (RAG). A featuring design choice of JusBuild is the “human-in-the-loop” approach, which allows the user (judge) to exercise his/her decision-making freedom and full control in the formulation of the provision in working with the suggestions provided by JusBuild. Thanks to the flexible nature of the architecture, adaptable to a large number of legal contexts, with different document structures and legal matters, JusBuild makes contextualized content generation accurate and efficient for legal practitioners. The application of JusBuild to legal document building in the Italian legal context is discussed. JusBuild validation is provided by considering datasets that differ for document template, language, and judicial matter, to test its applicability and adaptability to different contexts.