Building and deploying multi-agent RAG systems in 2024 with LlamaIndex

[@tonylofreso9579]

Agents starts @5:15

[@HM-gm1kn]

No matter what i do, create llama index doesnt work for me no matter how i set it up. So many errors and all the tutorials are outdated. Is there a discord or something I can go for help? I'm starting to feel like this build needs an update to function.

[@cagdasucar3932]

I don't understand why you would need a message queue between agents. Agents should not be micro services. It makes it slower and harder to debug.

[@jezieljones8508]

a lot of tasks take unspecified long amounts of time. client needs to be able to go off and do other things while waiting for the job to complete. the controlplaneserver abstraction does seem like a needless abstraction though

[@cagdasucar3932]

@@jezieljones8508 you can simply do an async call. You don't need micro services for that.

[@_nerdai_]

As these multi-agent systems become bigger and more complex, I think the sensible approach would be to break them apart into smaller components (i..e, good separation of concerns). In such situations the benefits would be something like monolithic apps vs. microservices. Also, probably not hard to imagine in the future where these multi-agent systems have some components performing compute/processing on data that cannot move, but maybe they still need to collaborate in some multi-agent setting. If you don't tease them out to their own service then I am not sure how else you would handle these situations. The microservice approach leads to more flexibility and increased scalability (as well as easier maintenance and debugging) when the systems get huge. If you're dealing with a small application where you don't expect to see huge amount of traffic or need many agents to collaborate, then I think you can go the monolithic approach and package the system in single service/docker.

[@_nerdai_]

@@jezieljones8508 Agreed. Yeah the most important function that the control plane does here is orchestrating (sub) tasks between the agents/human services. It doesn't do anything about scaling up the system in terms of number of replicas of an agent service however. That stuff is probably best handled with your container orchestrator like K8s.

[@cagdasucar3932]

@@_nerdai_ That's the theory but I think the reality is that you don't ever need such complex systems ever. I would challenge you to come up with a scenario that warrants micro services for agents. Premature optimization is the root of all evil, YAGNI and all that.

[@mihaelacostea5783]

Are you reading all your presentation from slides?? That's ridiculous.

[@vfryhn2]

Did you ever watch the video?

[@mihaelacostea5783]

@@vfryhn2 Yes, did you understand what I was saying? he is reading from some written text on his monitor, it sounds very artificial and i find it hard to follow.

[@vfryhn2]

@mihaelacostea5783 yes, you said "reading all your presentations from the slides" which if you watch the video you can see he is not reading "all" the presentations from the slides but using the slides to help present the information, there is a huge difference