Avoiding RAG Sprawl in the Enterprise
summary
Argues that enterprise RAG often produces fragmented, duplicative pipelines that break governance, security, cost and scaling; contrasts DIY stacks using LlamaIndex, LangChain and Hugging Face models with a single-platform approach (Vectara) that centralizes retrieval, reranking, prompt-guardrails, hallucination detection and correction. Shows a manufacturing app turning manuals into a phone‑based repair assistant, presents a hallucination leaderboard (Google Gemini 2.0 flash ~7% grounded rate) and claims a real‑time detector called mockingbird 0.9% with 3.8 million downloads on Hugging Face.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 6:42AI in enterprise future
- 13:07RAG maturity stages
- 17:39RAG sprawl explained
- 22:28Enterprise RAG benefits
- 26:23Responsible AI risks
- 29:50Hallucination measurement
- 32:41AI agents and hallucinations
- 33:03Q&A
related talks
31mAdvanced Enterprise RAG Systems
42mTo RAG or not to RAG
11mBuilding easily deployable RAG applications with Llamastack
19mThe RAG Stack We Landed On After 37 Fails
36mProduction Ready RAG
32mBuilding Multi-Agent RAG Systems
14mBeyond RAG: Building Reliable AI Systems for Privacy Assessments
48mCrafting your own RAG system: Leveraging 30+ LLMs for enhanced performance