A RAG system's errors often stem from the retrieval layer, not the generator. In one case, a support assistant's bot cited a real internal document but gave a wrong answer because the passage containing the correct answer was not retrieved; the model faithfully summarized the wrong context it was handed. The author notes that debugging focuses on the visible generation step—rewriting prompts—while retrieval, a vector similarity search returning chunk IDs, goes unmeasured. The article lists five retrieval failure modes, none fixed by prompt changes: 1) Naive fixed-size chunking, where splitting every 512 tokens severs facts from their context (e.g., "$4,200" without its meaning as a refund cap); 2) Embedding mismatch.