In deepset Cloud, a new hallucination detector for retrieval-augmented generation has been implemented to combat the issue of large language models (LLMs) generating false information, known as "hallucinations". The detector uses a fine-tuned DeBERTa model to compare an LLM's output to a ground truth and produces scores that indicate how semantically similar a sentence is to a given source document. This allows users to identify when a model is hallucinating and take steps to mitigate the issue, such as showing only fully supported sentences or including a disclaimer for partially supported claims. The detector has been trained on data from Ohio State University's project on evaluating attribution by large language models and has been shown to outperform other approaches, with scores that beat their best model by ten points.