RAG is a data problem wearing an AI costume

Teams spend weeks A/B testing models and minutes on their retrieval pipeline. Then they wonder why the answers are confidently wrong. The model was never the bottleneck.

Retrieval quality is a data-engineering problem: chunking that respects document structure, metadata that lets you filter before you rank, and evaluation that measures whether the right passage was even in the context window. Get those wrong and no model, however capable, can recover.

The fix is unglamorous. Instrument retrieval separately from generation, build a golden set of question/passage pairs, and treat 'did we fetch the right thing' as a first-class metric. Trust follows retrieval, not the other way around.