Why This Matters

You build a RAG pipeline. Retrieval works: the right chunk is definitely in the top results. You send everything to the LLM, and it still gives a wrong or incomplete answer, as if it never saw the one piece of text that contained the answer.

Very often …