RAG Hallucinations

In a production-level RAG pipeline, providing external context does not completely eliminate model hallucinations. Identify the primary failure points within the RAG lifecycle (Retrieval, Augmentation, and Generation) that lead to these inaccuracies. How would you design a multi-layered mitigation strategy to ensure the final output remains grounded in the provided context?