Why your "RAG works" demo doesn't hold in production

The RAG demo always works. You point the model at 50 PDFs, ask three questions, get three impressive answers, ship the system. Then real users hit it and 30% of answers are wrong.

What breaks

The demo dataset is curated. Real corpora have duplicates, near-duplicates, outdated revisions, and inconsistent vocabularies. Retrieval pulls the wrong chunk and the LLM confidently summarises it.

What fixes it

Hybrid retrieval (sparse + dense), reranking, evaluation against a golden set, and a model that says "I do not know" when retrieval confidence is low. None of these are demo-stage concerns — all are production-blockers.