An Adaptive RAG-Based Question-Answering System in the Context of Industry 5.0

Abstract

In this paper, a pragmatic literature review approach has been shown to select research papers to determine important technologies in the context of Industry 5.0 or I 5.0 such as Artificial Intelligence (AI), Internet of Things (IOT), Collaborative Robot (Cobot), Cyber Physical System (CPS), Human Machine Interface (HMI), Edge Computing, Big Data, Digital Twin, Virtual Reality (VR), Reinforcement Learning (RL), Large Language Model (LLM), Multiple Criteria Decision Analysis (MCDA) etc. The crux of this paper is to develop an economical Adaptive RAG-based (ARAG) system which could generate contextual relevant responses to a user’s query. A two-stage hybrid zero-resource hallucination detection system has been developed to detect hallucinations in the generated response. A binary classifier has been developed using Mistral 7B for fact checking against reliable resources and a panel of multiple Large Language Models (LLMs), namely, Mistral 7B, Llama 3 8B, and Llama 2 7B, has been used to evaluate the factual accuracy of the generated response asynchronously using the 5-point Agreement Scale. Mistral 7B has shown a very high correlation with human judges. Open source or free resources are used to develop the ARAG, and, thus, the ARAG is economical. A brief discussion on multilingual responses is also included.