RAG systems can confidently generate answers that contradict their own retrieved context, with no errors anywhere to flag it. This article builds a self-healing layer that validates retrieval quality before generation, verifies grounding after generation using a separate LLM-as-judge, retries once with a rewritten query on failure, and falls back to an honest "couldn't verify this" response otherwise. Every failure gets logged asynchronously to Kafka, where independent consumers automatically gr...