Seven AI Agents and a Knowledge-Graph: AGENTiGraph

[@ahmadzaimhilmi]

Got any github repo for this paper?

[@dennisestenson7820]

There's a link in section 5 of the paper, but only for a component of the knowledge graph system.

[@crippsverse]

The most exciting part of all this development in agents is being able to do it all locally with an open source model that doesn't require an expensive computer.

[@gamer-gw9iy]

This is seriously exciting work. Thanks for sharing. Hope your channel blows up with this high quality content you're pumping out 👀 maybe try changing the color or fonts or something see if you'll get more clicks

[@harikantipudi8668]

This is really good but how does this solution scale in real time production grade workflows. Some of these novel approaches will only add value when they are scaled and implemented for real use cases. The challenges with engineering these solutions are for real

[@MusingsAndIdeas]

This research doesn't seem to be so much about scale, but about experimentation with different mental architectures to start understanding the basic principles behind systems integration and design with these already highly complex systems

[@dragoon347]

Already working on this independently for network sec hehe

[@johnkintree763]

Agreed. Excellent synthesis of multiple papers. LightRAG may have solved mapping between entities and relationships with a hybrid of vector and graph databases. Lightweight for improved performance. Knowledge is inclusive of sentiment, evidence, and reason. Agentic workflows should include the human in the loop to detect and correct mistakes by the models. It feels like we are getting close to some kind of convergence of a Cambrian explosion in collective terrestrial intelligence.

[@IvarDaigon]

bringing the product to market and making sure it can scale is the job of companies and engineers, not universities and research scientists. They just tell us how to get good results and we implement them into our products if it can be done in a cost effective way and if the research lives up to the hype.

[@dragoon347]

@@IvarDaigon correct luckily we have everything set up wit Kubernetes

[@Fleniken]

What does 90% task execution mean? What kind of tasks?

[@drfabraz]

Thanks for brilliant explanation. You mentioned that you have a lot of physics specialized BERT. Can you expand on how do you build such systems? Thanks in advance

[@IvarDaigon]

I'm pretty sure that when they say "Use BERT" they arent implying that you roll your own BERT model, just use any of the plethora of BERT based vector embedding models that exist on hugging face. Using ANY of the existing general pre-trained embedding models should be good enough because you are only doing semantic matching, just pick the one that gives good benchmark scores for semantic matching. The only time you would need a custom model is when there is new terminology that was invented after the model was trained or somehow did not make it into the training data.

[@code4AI]

You never worked on a mathematical domain, or a biochemical domain or a medical domain? You really think: "Using ANY of the existing general pre-trained embedding models should be good enough"?

[@dragoon347]

Ohhh trust me

[@johnkintree763]

Background sounds from a child do not bother me.