MedRAG Nexus: An AI-Powered Health Intelligence System Using Retrieval-Augmented Generation and Agentic AI

Abstract: A significant and underappreciated challenge in modern healthcare is the communicative divide between clinical documentation and the patients those documents describe. Pathology reports, handwritten prescriptions, and physician summaries are routinely generated but rarely understood by the individuals who receive them — creating a measurable gap between information delivery and informed patient action. This paper presents MedRAG Nexus, a tri- layered AI-powered health intelligence platform designed to close this gap through a RAG-grounded conversational interface. The system processes clinical documents via a multimodal Vision-AI pipeline — employing a fine-tuned TrOCR model for handwritten prescription parsing and an EfficientNet-B7 network for dermatological anomaly classification — and anchors all clinical reasoning in a Retrieval-Augmented Generation (RAG) framework built on ChromaDB and LangChain. Patients interact with their own medical records through natural language; the chatbot retrieves verified clinical knowledge before generating every response, ensuring that answers are evidence-grounded rather than model-generated. A LangGraph-orchestrated ReAct agentic layer enables autonomous interventions — including specialist appointment scheduling via Google Calendar and patient notification via the Twilio WhatsApp API — when document analysis identifies clinically significant findings. The architecture directly addresses two dominant failure modes of current health AI: inaccessible medical documentation and factual hallucination in general-purpose large language models (LLMs). Experimental design targets a prescription parsing accuracy exceeding 92%, a hallucination rate approaching zero through retriever-grounding, and a demonstrable reduction in patient time-to-care. This paper details the full system architecture, data flow, model selection rationale, ethical safeguards, and a structured validation protocol suitable for clinical evaluation.

Keywords: Retrieval-Augmented Generation (RAG); Large Language Models; Clinical Document Understanding; TrOCR; EfficientNet; LangGraph; Agentic AI; Clinical NLP; ChromaDB; Health Informatics; ReAct Framework; Medical AI; Prescription Parsing; Patient Health Literacy.