Résumé:
In the context of increasing legal complexity and digital transformation, legal departments in large organizations face increasing demands for timely and accurate information. This thesis addresses the problem of improving access to legal and regulatory knowledge for employees at Djezzy, a major telecommunications company in Algeria. The main objective is to develop an intelligent chatbot system capable of answering legal questions in French, with a focus on compliance, data protection, and internal policies.
To achieve this, we designed a Retrieval-Augmented Generation (RAG) system that com- bines Semantic Document Retrieval and Natural Language Generation using Large Lan- guage Models (LLMs). The architecture is built around a pipeline that indexes legal texts into a vector database and uses local LLMs via Ollama to generate responses based on retrieved content.
The evaluation was, successfully, carried out using automatic and human-based methods. We used semantic similarity, ROUGE lexical metrics, BLEU, and BERTScore to measure response quality and contextual relevance. In addition, a Djezzy legal expert reviewed the output, and large language models were also used to simulate judgement during evaluation.
The chatbot was implemented in a web application called JuriBot. This work demon- strates the viability of integrating RAG-based systems into enterprise legal services, of- fering a valuable tool for improving compliance efficiency and access to legal information.
Keywords: Legal Chatbot; Legal Assistant; RAG Technique, Large Language Models, Vector Search.
