Craig Anderton
Well-known member
Disclaimer: I don't know a thing about software and coding. But I am inherently curious, and I wondered if Studio Assistant is an example of a small, trained LLM. Well, I've been digging into how this works, and it's very interesting. I think Studio Pro is basically trying to end Help files, and looking into an assistance mechanism that may not be future-proof, but sure seems to be future-friendly.
Studio Assistant acts like you're interacting with ChatGPT or its ilk, but it's based on a generic, off-the-shelf LLM from Mendix. It's not trained on Studio One, it's just a generic LLM. So how does it "learn" Studio Pro? The answer is...it doesn't. Mendix is embedded within Studio Pro, and basically wraps the off-the-shelf LLM with Mendix-specific guardrails and tools. So, the LLM isn't trained on Studio Pro, or becomes an "expert" on how Studio Pro works. Instead, it's an agent that does information retrieval. Because it's inside Studio Pro, it knows what document you're looking at, what you're trying to do, and the context of what you want to know. It can retrieve documentation, examples, metadata - basically, everything that exists in the Studio One environment. This is why it doesn't have to "learn" anything, it just finds relevant stuff that already exists.
This seems like a really smart approach. It's the same type of architecture used by Microsoft Copilot, Google Gemini Agents, OpenAI GPTs, etc., applied specifically to our world. Mendix doesn't have to be retrained when the next update or version of Studio Pro comes out, nor does it need a huge proprietary dataset. There are very low hallucination risks, and no legal/IP complications because it's interacting in a strictly regulated environment with guardrails. Software geeks who know more about this stuff than I ever will can check out this link about Mendix AI Assistance. And please, correct me if I'm wrong and missing or misinterpreting something! Although I'm concerned that this will become a "Studio Pro subscribers get X amount of free tokens and then you have to pay," to be fair, not everything can be free.
As Spock used to say, "Fascinating, captain." Or at least I think so
Studio Assistant acts like you're interacting with ChatGPT or its ilk, but it's based on a generic, off-the-shelf LLM from Mendix. It's not trained on Studio One, it's just a generic LLM. So how does it "learn" Studio Pro? The answer is...it doesn't. Mendix is embedded within Studio Pro, and basically wraps the off-the-shelf LLM with Mendix-specific guardrails and tools. So, the LLM isn't trained on Studio Pro, or becomes an "expert" on how Studio Pro works. Instead, it's an agent that does information retrieval. Because it's inside Studio Pro, it knows what document you're looking at, what you're trying to do, and the context of what you want to know. It can retrieve documentation, examples, metadata - basically, everything that exists in the Studio One environment. This is why it doesn't have to "learn" anything, it just finds relevant stuff that already exists.
This seems like a really smart approach. It's the same type of architecture used by Microsoft Copilot, Google Gemini Agents, OpenAI GPTs, etc., applied specifically to our world. Mendix doesn't have to be retrained when the next update or version of Studio Pro comes out, nor does it need a huge proprietary dataset. There are very low hallucination risks, and no legal/IP complications because it's interacting in a strictly regulated environment with guardrails. Software geeks who know more about this stuff than I ever will can check out this link about Mendix AI Assistance. And please, correct me if I'm wrong and missing or misinterpreting something! Although I'm concerned that this will become a "Studio Pro subscribers get X amount of free tokens and then you have to pay," to be fair, not everything can be free.
As Spock used to say, "Fascinating, captain." Or at least I think so
