Naive RAG seems simple: chunk docs → embed → retrieve top-K → send to LLM. But the cracks show fast: → Fixed chunks split context mid-thought → Top-K retrieval has no idea what's truly relevant — just what's similar → No query rewriting = garbage in, garbage out

→ Hallucinations slip through when retrieved chunks conflict → Zero re-ranking means noisy context crowds out the good stuff Retrieve smarter, not bigger.