One of my biggest motivations for writing about AI is my deep belief that AI must be UNDERSTOOD.
In the first part of this series, I introduced AI Assistants; then, I tried to understand their diet to explain it.
Now, let's go into how they work because we need to understand AI to deal with it with the right approach.
By the way, the homework for today was to think about whether the first small project could be a public or private Assistant (i.e., just for your organization) and powered by public or confidential data.
Have you done them?
Have you chosen to start with a Private Assistant based on Public Data?
Well, it's the right choice. The other combinations all have much higher degrees of difficulty and risk.
And now, let's try to understand how they are made.
What's behind the scenes?
2023 was the year that saw hundreds of “ready-made solutions” dedicated to interacting with your documents or creating these types of solutions.
Even if you are not a technician, understanding how they are made can help you clarify the results you can obtain.
For example, many confuse AI with search engines (Not that the companies that develop both are trying to tell you anything different). But things are other: Generative AI DOES NOT SEARCH in an archive but only tries to predict the best next word (Here, you will find a post from some time ago on this topic).
In any case, the best solutions combine different technologies such as Generative AI, traditional software (such as Chat interfaces), conventional or vector databases (which, explained, serve to store data in a way accessible to algorithms), APIs (i.e., interfaces which are used to make the software talk to each other) and mixed techniques such as RAG (Retrieval Augmented Generation) which I have already talked about somewhere on the blog.
And now, blue pill or red pill?
🔵 Blue pill: if you don't want to suffer over the technicalities, stop here; I'm not offended. You can skip the next paragraph without any problems. Click here 🙂
🔴 Red pill: enter a little into the technical field and, before making decisions, understand how the architectures that will allow your company to make the best use of Generative AI are structured. Considering that things change quickly, you must grasp the change well.
🔴 What does an AI Assistant look like behind the scenes?
You chose the red pill, good!
Below, you will find a typical architecture of an AI Assistant capable of using your documents and data in the various databases and the selected Generative AI model.
Each solution is different and could differ. But this representation gives a reasonably general and clear idea of their work. Then, as always, it can be as complicated as you like in computer science.
A typical architecture of a Generative AI-based Chatbot solution with RAG technology. The parts in BLUE are 'normal' software. In purple are those most linked to the new worlds of AI. The ones in orange are your data, both in input and output.
If you don't like the videos, click here...
But do they just read and generate text?
No!
I didn't include it in the graph, but I wouldn't want to dwell on it here because this topic deserves all its space. AI Assistants are already evolving into AI Agents. That is, software that exploits the conversational dynamics of LLMs performs actions and automation on the traditional (and non-traditional) software systems on which they can operate.
There are various levels of autonomy in the decisions these Agents can make. The first level, the easiest, is the one that requires it to perform pre-configured actions. In GPT Chat, they are called Actions, and I talk about them in the second chapter on GPTs
Already with this first level of autonomy, an assistant could:
Automatically open a ticket on your helpdesk like in this example we are using in greenride.it :
2. Perform tasks on your ERP (Create documents, run workflows)
3. Produce graphic content (for instance, on Canva or in PowerPoint)
4. Publish content on social media
5. Perform actions built on automation engines like n8n or Zapier like the following, a workflow manageable by an LLM-based Agent extracted from VJAL INSTITUTE's CEO EMPOWERMENT workshop.
and… every thousand software automation tasks you can think of performing starting from a prompt.
Now that we understand a little more that we are faced with a solution composed of traditional software and LLM models comes the important question:
Do we buy it or do it?
The architectures can vary greatly and are not always made explicit by the suppliers who offer them. Choosing whether to start with a homemade solution or a ready-made one depends on many factors.
You essentially have these possibilities:
Adopt Turnkey solutions made for this purpose by large or small emerging companies.
Build a custom solution.
We'll talk about it in the next episode. I won't give you homework but I advise you to come prepared by reviewing all the previous posts.
🚀 To always keep you updated on my contents:
📝 Subscribe to the Glimpse Blog to never miss an update
📚 Check out 'Glimpse', my novel about artificial intelligence
🗓️ Contact me if you want to organize an AI Workshop or for any idea.
See you soon!
Maximilian
Comments