IDENTIFYING HULLCINATION IN RETREIVAL AUGMENTED GENERATION

Abstract

Large Language Models (LLMs) have been applied to build several automation and personalized question-answering prototypes so far by blending the strengths of Large Language Models (LLMs) with the dynamic retrieval of external documents to enhance the quality and relevance of generated content. However, scaling such prototypes to robust products with minimized hallucinations or fake responses still remains an open challenge, especially in niche data heavy domains such as medical or financial industry. Understanding and diagnosing these hallucinations is crucial. It not only ensures the integrity and reliability of the information provided by these models but also pushes the boundaries of AI’s utility in critical applications. The ongoing problem of hallucinations has tempered enterprise enthusiasm for generative artificial intelligence (AI). While large language models (LLMs) can generate rapid, confident and fluent responses to many user prompts, some are demonstrably false. Be- cause of the closed nature of these systems, systematically assessing these claims is challenging and explianing those responses are much harder. In this work, we discussed an approach to detect the hallucination in RAG responses caused by the retrieved information and provide the metric of hallucination in these response. This work give a practitioner to detect hallucinations during the runtime and mitigate those responses to enhance the reliability and relevance of the generated responses from RAG systems. Particularly, we focus on identifying the root cause of hallucinations scoping are study on retrieved information and generated response.