There’s been a bunch of questions thrown up about Gemini 1.5’s implications for RAG. I’m pretty bullish that Gemini 1.5 will accelerate RAG systems to production. Gemini 1.5’s 40% failure rate in multiple-needle recall, as well as lack of meaningful ways to debug the black-box of an LLM will mean more auditable systems like RAG will give back control to developers. Further, at end-state, being able to optimize for Accuracy, Costs and Latency are also issues that mean that the last-mile of optimization will always be RAG-based. Curious to learn more about how others think! https://medium.com/enterprise-rag/why-gemini-1-5-and-other-large-context-models-are-bullish-for-rag-ce3218930bb4

I'm also interested in ring attention stuff here for larger context lengths, which is going to be huge for RAG: