ReRag: A New Architecture for Reducing the Hallucination by Retrieval- Augmented Generation

As the application areas of Generative AI models become widespread, new problems arise. One of these problems is the “hallucination” observed in Large Language Models (LLM). Recently, Retrieval-Augmented Generation systems (RAG) have been introduced to cope with this problem. In this paper, we propose a new method, called “ReRag for Retrieval Augmented Generation”, to reduce the hallucination effect of LLM responses. This method assumes that a small dataset of queries and their man-made ideal responses are available for a specific application domain. It attempts to optimize the hyperparameters of the vector database in the RAG system by generating close-to-ideal responses concerning this ideal dataset, in which the questions and the ideal answers are provided. ReRag measures the similarity between the generated and ideal responses and then modifies the hyperparameters iteratively to increase the similarity between these two responses. Experimental results show that the ReRag method reduces the hallucination problem and generates relatively close to ideal responses compared to raw models.