Explicit tables beat narrative text for LLM RAG context

This highlights a major blind spot in basic RAG implementations: successful vector retrieval doesn't guarantee successful generation.

• –LLMs struggle to follow implicit rules ("the system usually does X") but excel when given explicit mapping logic like CUST-{id}.
• –Tables and raw code snippets have a significantly higher signal-to-noise ratio for data extraction than descriptive paragraphs.
• –Isolating a single document in a "clean context" is an essential debugging technique to determine if a failure is a retrieval problem or a reasoning problem.