LegisSearch: navigating legislation with graphs and large language models

Navigating and retrieving relevant excerpts of legislation is challenging, requiring time and effort, especially to fine-tune appropriate input search queries. Furthermore, the continuously growing, heterogeneous body of laws, combined with a deep interconnection among normative acts, adds a layer of complexity: some potentially relevant rules may be hidden in articles that, through multiple citations and references, might be relevant for the input query. Traditional search systems, based on keywords or more sophisticated approaches as BM25 or TF-IDF, do not support such flexible exploration, being ineffective at handling contextual information. To address these challenges, recent research proposed using graph data models for legislative knowledge management, introducing a straightforward approach to handling network complexity. They adopted the Property Graph data structure, demonstrating how it provides semantics and navigation power, supporting advanced querying tools for legislative acts, and implemented it on the Italian legislation. In this paper, we build on recent results on legislative knowledge management with graphs by proposing LegisSearch, an effective navigation system that, combining the graph data model with pre-trained Large Language Models and universal text embeddings, allows users to conduct powerful searches within a legislative system. We implement LegisSearch within the Italian graph of national laws, and we test its performance across multiple domains by comparing its search results with those provided in specific thematic areas by Italian ministries on their official websites, demonstrating its superior retrieval performance over traditional search systems and testing the contribution of each component.