RAG is dead: why enterprises are shifting to agent-based AI architectures
A colleague asked me about this recently:
I don’t think that this article supports the assertion that “RAG is dead”. It’s simply pointing out that Agentic AI can be tasked with fetching the most current versions of the organisation's documents representing the enterprise's orthodoxy. RAG still represents that orthodoxy. What he's selling is using Agentic AI to ensure the freshest versions of corpora that are in the RAG collection for a given enterprise project/task/activity are used.
IMO, Some of the author's assertions are "word salad" and some are a bit of a stretch (shall we say). For instance, the scaling and scoping problem--No one in their right mind believes that all the organisation's orthodoxy should be concentrated into a single impossible-to-manage repository; that's simply not the way governments or enterprises operate. And there are a number of other assertions I could take issue with, but none of them invalidate RAG as one of the many tools to assure good generations from LLMs.
I defer to our esteemed department Chair for his learned opinion, but again IMO, this is a sales and marketing ploy, which we are led to by a rather "click-bait-ish" Headline.