Abstract: Eye diseases such as conjunctivitis, glaucoma, cataract, diabetic retinopathy, and optic neuritis are becoming increasingly common worldwide. If not detected at an early stage, these conditions can lead to serious vision impairment or even permanent blindness. In many parts of the world, especially in rural and remote areas, people face significant difficulty in accessing a qualified eye specialist on time. This paper presents a Confidence-Guided Multi-Agent Large Language Model (LLM) Framework developed for clinical decision support in ophthalmology. The system accepts symptom descriptions from patients written in plain English and processes them through a pipeline of five intelligent agents. Each agent performs a specific role, including symptom analysis, disease retrieval using Retrieval-Augmented Generation (RAG) with FAISS, web-based validation using DuckDuckGo, and final decision synthesis. A unique confidence scoring mechanism is incorporated, which combines the RAG similarity score and the web trust score to give users a reliable percentage-based prediction. The system is built using Python and Streamlit, making it accessible through any standard web browser without requiring specialized software. Experimental results demonstrate that the system correctly identifies common eye diseases with confidence scores above 80% for valid inputs, and appropriately rejects unrelated inputs. This system is intended for academic demonstration and can serve as a supportive tool for telemedicine platforms and rural healthcare centers.

Downloads: | DOI: 10.17148/IJIREEICE.2026.14446