Retrieval-augmented generation is now considered the most effective way to ground a model in factual data by 80 percent of enterprise software developers, and roughly two-thirds of Fortune 500 companies are piloting a RAG-based internal knowledge base. But production numbers still tell a more complicated story: real-world enterprise RAG systems hallucinate on more than 10 percent of queries, with legal and medical domains regularly pushing past 20 percent, and accuracy that looks strong on single-hop factual questions can collapse to around 60 percent once a query requires reconciling conflicting sources. This session is about closing the gap between the RAG pitch and the RAG reality, specifically what breaks retrieval quality in production and what actually fixes it. What this session will cover: - Why RAG systems that score well on standard benchmarks still hallucinate at meaningfully higher rates on real enterprise queries - What happens to retrieval accuracy once a question requires reconciling multiple, possibly conflicting sources - Which domains, legal, medical and similarly high stakes areas, need a materially higher bar for grounding before going to production - What a realistic evaluation pipeline for a production RAG system actually measures, beyond a single accuracy number