The article examines the issue of hallucinations in large language models (LLMs), which are instances where these AI systems fabricate information, leading to significant challenges in deploying them in sensitive applications. Datadog has developed a real-time hallucination detection feature, especially for retrieval-augmented generation (RAG) scenarios, focusing on faithfulness—ensuring LLM-generated answers align with a given context. The company employs black-box detection methods, particularly LLM-as-a-judge approaches, to evaluate the accuracy of LLM outputs without accessing the model's internal workings. This involves a structured prompting strategy that breaks down tasks into smaller guided steps, improving accuracy by leveraging the LLM's strengths in guided summarization. Datadog's technique has shown promising results, particularly in challenging human-labeled benchmarks, and highlights the significant impact of prompt design over just model architecture in detecting hallucinations effectively. The company continues to refine its approach and invites interested individuals to join their team.