Advancing Multilingual Retrieval-Augmented Generation for Reliable Medication Counseling

Abstract

Recent advances in Large Language Models (LLMs) have opened opportunities for intelligent medical chatbots. However, most existing systems remain limited to English-language corpora and lack adaptation to regional drug repositories. This paper introduces a Retrieval-Augmented Generation (RAG) chatbot designed for medication counseling across Franco-Moroccan contexts. Our pipeline integrates 5,300 Moroccan and 8,300 French drugs into a unified knowledge base, indexed with SentenceTransformers and FAISS, and paired with state-of-the-art LLMs. Evaluation was conducted on a curated benchmark of 100 drug-related queries covering dosage, contraindications, and drug-drug interactions. Results show that our chatbot achieves a BERTScore F1 of 0.71. This corresponds to a +5% improvement in factuality, measured using automated faithfulness and groundedness metrics, compared to state-of-the-art baselines. Beyond numerical gains, the system demonstrates robustness in handling multilingual queries, contextual terminology, and region-specific prescriptions, addressing a gap in current healthcare AI. These results demonstrate the feasibility of deploying RAG-based systems in underrepresented linguistic and cultural settings, offering safer and more transparent medication information. This work advances the field of medical LLMs by combining semantic accuracy with contextual reliability, setting a precedent for multilingual and localized healthcare AI systems.