Development of a Retrieval-Enhanced Chatbot for the Italian Workplace Safety Regulation in Low-Resource Settings

Hallucination remains a major challenge when applying large pretrained language models to specialized domains such as legal or regulatory text. In contexts like Italian workplace safety regulations, unreliable or unsupported responses can lead to serious consequences. To address this issue while maintaining efficiency in low-resource environments, this thesis presents the design, implementation, and evaluation of a retrieval-enhanced chatbot tailored to Italian workplace safety regulations, optimized for low-resource deployment scenarios. The proposed system couples a compact pretrained language model with a document retriever so that generated responses are grounded in official regulatory texts and can include explicit citations. To validate the idea, we compare the chatbot's performance with and without the retriever using multiple-choice question benchmarks derived from workplace safety regulation materials. The results indicate that, while language models contain some latent knowledge of the domain, the retrieval-augmented configuration consistently yields higher accuracy and more reliability. We also attempt to fine-tune the retriever to improve recall; however, the off-the-shelf retriever is already strong, and fine-tuning produces negligible gains. These findings confirm that retrieval augmentation can substantially improve the reliability and accuracy of language models in specialized, low-resource regulatory domains.