A Document-Grounded RAG Chatbot for Medical Radiation QA : Improving Accuracy with HyDE, Caching, Multilingual Matching, and Cloud Deployment

Key Points

• The study aims to enhance the accuracy and efficiency of a chatbot providing QA for medical radiation equipment using document-grounded methods.
• Developed a chatbot based on retrieval-augmented generation (RAG) techniques.
• Applied query expansion and embedding techniques to generate virtual responses.
• Normalized responses into English to reduce inconsistencies in multilingual environments.
• Combined memory and file caching to lower response delays and model call costs.
• Conducted a standardized validation process with 120 evaluation queries and expert reviewers.
• Average accuracy score improved from 3.66 to 4.83.
• Search failure rate reduced from over 50% to 13.3%.
• Average response time decreased from 3.5 seconds to 1.8 seconds.
• Quality of responses evaluated on six metrics: accuracy, fidelity, relevance, consistency, expressiveness, completeness.

Abstract