The article examines evaluation models designed to automatically detect hallucinations in Retrieval-Augmented Generation (RAG) systems and benchmarks their performance across six RAG applications. RAG systems enhance AI by incorporating company-specific knowledge, thereby reducing but not eliminating hallucinations, which remain a significant issue affecting trust and usability. The evaluation models assessed include LLM-as-a-judge, Hughes Hallucination Evaluation Model (HHEM), Prometheus, Patronus Lynx, and Trustworthy Language Model (TLM), each offering different approaches to assess response accuracy without relying on ground-truth answers. The study's benchmark methodology focuses on the models' ability to flag incorrect responses using precision and recall metrics, and it reports that models like TLM and LLM-as-a-judge often outperform others in detecting inaccuracies, particularly in datasets such as FinQA and ELI5. Despite some models being specially trained for specific errors, the study suggests that general-purpose models like TLM may remain more adaptable to future LLM variations. Additionally, the article highlights the importance of choosing the right evaluation model based on the specific domain and dataset characteristics.