Mitigating AI Hallucinations: Rag as Support for Safe Antimicrobial Prescribing in the Era of Artificial Intelligence

Antimicrobial resistance (AMR) is a global health emergency, often fueled by inappropriate antibiotic use. In this context, this study investigates the potential of Artificial Intelligence (AI) to improve clinical decision-making by evaluating the effectiveness of a model known as Retrieval-Augmented Generation (RAG) applied to antimicrobial prescribing by non-infectious disease physicians. The focus is on a moderation tool designed to operate over a commercially available, free-access generative AI (Gemini 2.0), aiming to reduce hallucinations ‒ plausible-sounding but technically incorrect answers. Such failures, common in isolated, unmoderated Large Language Models (LLMs), undermine the reliability of generated clinical information and represent a critical barrier to safe adoption in medical practice. Three moderation workflows were tested to steer the LLM toward evidence-based, concise, and source-validated answers: the WHO AWaRe antibiotic book. A conversational agent interface was developed, and three models were compared: (I) LLM without RAG (control), (II) RAG with Direct Vector Retrieval (RAG-Direct), and (III) RAG with Multi-Perspective Optimized Retrieval (RAG-MP). Evaluation used 50 challenge questions grounded in real clinical scenarios, and responses were assessed using the RAGAs framework, which measures accuracy, faithfulness, and retrieval effectiveness. RAG substantially increased the mean correctness of responses from 0.447 in the control model to 0.763 (RAG-Direct) and 0.855 (RAG-MP). RAG-MP achieved the highest mean accuracy and contextual recall (0.969), indicating more complete answers. RAG-Direct showed greater consistency, with a lower standard deviation in correctness (SD = 0.097), reflecting greater response stability. The RAG architecture proved to be an effective, accessible, and flexible solution to promote informational safety in antimicrobial prescriptions generated by AI. Choosing between RAG-Direct and RAG-MP requires balancing stability versus maximum accuracy, with both options technically viable and adaptable to the epidemiological and structural realities of different health services.