Operating production RAG at platform scale under continuous change

Enterprise RAG systems face continuous documentation updates, strict SLAs, and version-sensitive correctness, making them infrastructure challenges, not demos. This talk presents a production architecture deployed for Adobe’s AI Assistant, operating over 121K+ enterprise documents with version-aware retrieval. The speaker will cover hybrid sparse+dense retrieval achieving 72.8% nDCG@4 with sub-200 ms P95 latency, scalable generation of 700K+ closed-domain QA pairs, and regression-safe evaluation pipelines combining automated and human review to prevent retrieval and answer-quality failures in production.