Table of Contents
Since LLM-based applications are relatively new, it can sometimes be overwhelming to keep up with the marketing jibberish terms thrown at us.
For example, Microsoft released an article called "Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks".
Does this mean you need Magentic-One? Maybe not. This is what we want to explore in this article.
Note: This article aims at simplifying the concepts to make it accessible to people with less technical knowledge.
Step 1: AI vs LLM - the basics
If you are new to the topic, you might wonder about the difference between AI and LLM. To make this simple, LLMs (or Large Language Models) are a piece of software that uses AI to perform predictions.
ChatGPT is one of such LLMs. It uses an AI model to answer your question in a chat-like maner. As you have probably noticed, ChatGPT can answer in many different domains. Thats why the underlying model (GPT4o at the time of writing) is called a general (or generic) model.
But what about the word GenAI?
GenAI stands for Generative AI. It means that the tool, given a prompt will generate something. With tools like Midjourney, this "something" will be an image, with ChatGPT it will be text.
Step 2: RAG - the importance of the context
While LLMs are very good at giving you facts about general knowledge, they are unable to give you facts about how to solve a problem specific to your company.
That is because your company's know how is outside the LLMs context.
A context is the information available to the LLM to generate an answer. Since your company's know-how is very important, you probably want to keep it a secret.
But what if we could give the LLM relevant data to tell your employees or colleague how to deal with that annoying administrative form without having to train a new model? Wouldn't it be great?
Well that is the purpose of a RAG. RAG stands for Retrieval Augmented Generation. It uses your company's database and fetches relevant info before passing it to an LLM as context.
To be able to do this, your database should be of a specific type, namely a vectorial database. We won't explain this here but just know that they work differently than a regular relational database.
So to recap, a RAG is a piece of software that connects a database to a LLM in order to generate something (like text) that could not be generated with the AI's training data.
Step 3: Agents - interaction
So far we have been explaining that LLMs can generate something (mainly text) based on a prompt and a context. You can even ask an LLM to create an action plan on how to do something.
The third step is to allow your software to interact with its environment.
But first - what does to interact mean in the context of software?
To interact means to operate a change in the environment. For example, adding a new contact to a contact list, alerting about an appointment that is going to start, updating a phone number or cancelling a delivery are all interactions.
The second question is how?
Let me use a metaphor to explain this. Your sink is clogged and you call a plumber. The plumber comes to your home and takes a look at your sink. Then he reaches for his toolbox and chooses one of his tools to detach the pipes. Then he takes a second tool to clean the pipes. And finally he tightens the pipes.
Since an LLM is capable of creating a plan of action, we use the LLM as the brains of the operation. It is the plumber.
Just like the plumber, the LLM needs a toolbox. The tools are regular pieces of software (functions, methods, applications,...)
The LLM will look at the problem and design a plan of action that uses the tools in its toolbox.
We will give the LLM the rights to execute software within the limits of the tools.
In the same way, your plumber wouldn't be able to trim your hedge, an LLM cannot do something its toolbox isn't designed for.
An application or system that is agentic is an application or system that uses an agent.
Step 4: Multi-agent - orchestrating agents
Based on what we've seen above, you might have guessed that a multi-agent system is a piece of software that makes use of multiple agents.
The important notion to remember is that multi-agent does not mean to have multiple instances of the same agent but rather having a set of different agents that are each responsible for different things.
In non-technical terms, this means that your company is not made up of multiple plumbers but rather of plumbers, gardeners, woodworkers and accountants.
When we are looking at it from a human perspective, it makes sense that there must be a manger or CEO that oversees the workload and sets the priorities.
In the technical world, this is called an orchestrator.
Conclusion
Together, we have seen a high level definition of multiple concepts used in the GenAI industry.
We have given a brief definition of LLMs and GenAI. We have described RAGs and Agents and have introduced the concept of multi-agent systems.
This should allow you to better understand which type of infrastructure your project needs.
