LegalDefAgent

The complexity, dynamicity over time, and multilingual nature of legislative documents pose significant challenges for the accurate retrieval and reuse of legislative definitions, an essential task in legal drafting.

LegalDefAgent an AI-driven system leveraging Large Language Models (LLMs) to assist in the retrieval and generation of legal definitions from a multilingual, multi-jurisdictional dataset of XML-encoded legislative documents.

The system functions as a conversational AI agent, enabling natural language queries tailored to different end-user types, such as lawyers, legislators, and judges.

It employs a hybrid retrieval approach, integrating dense semantic search with sparse keyword-based methods, and incorporates legislation-aware and point-in-time filtering to ensure jurisdictional and temporal accuracy. If no suitable definition is found, the system leverages Retrieval-Augmented Generation (RAG) to generate a novel one that is grounded in and consistent with in-force legislative documents.

The system is evaluated using automatic quantitative metrics and qualitative assessments from legal experts, demonstrating strong retrieval capabilities but highlighting limitations in generating legally sound definitions.

Installation

# Clone the repository
git clone https://github.com/leonardozilli/LegalDefAgent

# Move to the repository folder
cd LegalDefAgent

# Install package and required dependencies using uv
# uv install options: https://docs.astral.sh/uv/getting-started/installation/
uv sync

# Activate the virtual environment
source .venv/bin/activate

Configuration

Copy and rename the environment template .env-example to .env and populate it with the required credentials.

Usage

LegalDefAgent provides a CLI for common operations. You can invoke it with:

legaldefagent [COMMAND] <args>

Available commands:

• extract-definitions : Extract definitions from local XML files or eXistDB collections.
• embed-definitions : Compute embeddings for extracted definitions.
• populate-vectorstore: Populate the vector store with the generated embeddings and metadata.
• run-service : Start the backend agent service.
• run-app : Start the Streamlit frontend app.

Example workflow:

# Extract and embed definitions
legaldefagent extract-definitions -s exist

legaldefagent embed-definitions -i data/definitions_corpus/definitions.csv

# Populate the vector store
legaldefagent populate-vectorstore -d data/definitions_corpus/definitions.csv -e data/embeddings/defs_embeddings_hybrid.pkl

# Start the FastAPI server
python -m legaldefagent.cli run-service

# In a separate terminal, launch the Streamlit app
python -m legaldefagent.cli run-app

Docker Setup

Alternatively, run the application using Docker:

1. Build base image:

   docker build -t legaldefagent-base:latest -f Dockerfile.base .

2. Build and start the services:

   docker-compose up --build

The services will be available at:

• FastAPI: http://localhost:8000
• Streamlit: http://localhost:3000

Documentation

Architecture Overview

• Milvus stores embeddings of legal definitions.

• eXist-db stores XML sources (Akoma Ntoso, EurLex, Normattiva).

• API bridges queries between frontend, embeddings, and XML collections.

Definition Retrieval Pipeline

License

This project is licensed under the MIT License - see the LICENSE file for details.