Implementation of Retrieval Augmented Generation in a Web-Based Dermatological Chatbot System

Abstract

Indonesia’s tropical climate, poor sanitation, and limited access to medical services especially in remote areas are key factors contributing to the high prevalence of skin diseases. Direct access to dermatologists remains difficult for many people. This study aims to develop a dermatological consultation Chatbot using a Retrieval Augmented Generation (RAG) approach, leveraging the LangChain framework, the LLaMA model, and the Qdrant vector database. The dataset includes 30 types of skin diseases sourced from the National Library of Medicine. The preprocessing stage involved whitespace normalization, removal of special characters, and handling of missing values to ensure data consistency before vectorization. Evaluation results showed high scores for Faithfulness (0.9429) and LLMContextRecall (0.9600), indicating that the responses were relevant and aligned with the source documents. However, a relatively low Precision score (0.4720) suggests a need for improved information accuracy. The Chatbot is integrated with the Chainlit platform, offering an interactive user interface that supports login, conversation history, and user feedback features to facilitate system development based on user input. The system demonstrated fast retrieval times (0.08–0.29 seconds), though answer generation remains slow due to CPU infrastructure limitations (255–283 seconds). Future improvements should focus on enhancing answer accuracy, optimizing the model's performance, enriching the medical reference dataset, and adding automated medical validation features to ensure the reliability of consultations. Therefore, this Chatbot system is expected to serve as a cost-effective and efficient alternative for providing initial information on skin conditions to individuals with limited access to healthcare services.