The application of generative AI in highly regulated and specialized areas, such as Italian agricultural law, is associated with a number of challenges related to data confidentiality, domain specificity, and hallucinations. In order to address these challenges, we designed a Retrieval-Augmented Generation (RAG) model that is strictly grounded in authoritative legislative sources. The proposed system uses a hybrid dense-sparse retrieval approach on a curated corpus of Italian legislative texts, carefully segmented at the article level. The performance of two compact models, Gemma and Granite, was tested using a dataset of real-world legal questions obtained from professional exams, with a focus on legal grounding. The experimental results show that, when combined with accurate retrieval, Small Language Models (SLMs) are able to provide reliable and contextually grounded legal reasoning. These results highlight both the feasibility and the current limitations of using small-scale models for domain-specific LegalAI applications in civil law systems.