Ernesto Amato

Spreading innovation and technology with passion and enthusiasm!

Basics

Here you can find some simple Q&A

  • AI Assistant vs RAG vs AI Agent

    We really need some clarifications. Nowadays I’m hearing a lot of people talking about AI Agents, the newest trend in AI.

    Well, most of the times, when I ask to explain what they mean with ‘agent’ the answer is typically the… wrong one!

    And it gets worse when we talk about RAGs.

    First things to know: all of them are based on foundation AI models and use GenAI models to interact with users (if you still have some doubts about what GenAI is you can read this article: Generative AI explained, in a (very) simple way).

    Ok, let’s try to explain what they are..

    AI assistant

    “It is a software application that uses artificial intelligence to understand user commands, process them, and perform specific tasks to help users.”

    This definition is created by Gemini.

    Some early examples of AI Assistants have been Amazon’ s Alexa, Apple’s Siri, Google’s Assistant, etc.

    Today, thanks to GenAI models, you can interact with AI Assistants using NLP (Natural Language Processing) and you can personalize them providing detailed descriptions of the context in which the assistant could work and specific information (via files or links or other sources) to be used as a knowledge base.

    Note: these “additional” knowledge bases are not used to train the model.

    Now you can start building your AI Assistant by describing what its purpose is, how to use files within the knowledge base, how to connect via API to external services, etc. Then, you add your prompt, and the AI Assistant will answer with the right information based on what you provided using its foundation model… hopefully 🙂

    RAG (or KAG or CAG or…)

    Retrieval Augmented Generation (RAG) is a technic to optimize LLMs providing direct connection to external knowledge bases.

    As you know, LLMs are trained using large data sets, but most of the time they have generic knowledge of many topics.

    When you need to train one LLM to a domain-specific topic then you can use RAGs.

    A RAG first retrieves relevant information from external sources and then uses that information to generate the answer.

    You can develop Enterprise RAGs to offer to your employees a specific knowledge-based AI Assistant.

    If AI Assistant is an application, RAG could be the underlying technique used to make “plain” assistants more knowledgeable and reliable.

    AI Agent

    I like this definition given by ibm.com web site: “If AI assistants are reactive, performing tasks at your request, AI Agents are proactive, working autonomously to achieve a specific goal by any means at their disposal.“

    Applications based on AI Agents (have you heard about Agentic Commerce?) seem to be the new buzzword in the Generative AI field. Maybe the real “killer application” that will finally provide a real ROI.

    We are talking about decision-making capabilities.

    One of the best examples available today is the “AI web browser” field: Comet by Perplexity and Project Mariner by Google DeepMind.

    Both promise to offer a highly personalized experience (if you open your life information to them…) and decision-making capabilities. You can also use them to fill web based forms, maybe examination forms (https://fortune.com/2025/10/10/ai-cheating-on-homework-chatbots-students-education-perplexity-ceo-coursera/)

    Today, you can also create AI agents in a simple way thanks to Gemini 2.5 Pro Computer Use feature.

    Conclusions

    That’s it! Now you know a little bit more about GenAI systems. I’m sure you also understood that you should use AI Assistants (using RAGs) or AI Agents based on your specific needs… and their costs.

  • Generative AI explained, in a (very) simple way

    Novembre 30th 2022: OpenAI revels it’s LLM, ChatGPT! One of the first examples of Generative AI. But, exactly, What GenAI is?

    Let’s start moving some step back to types of Artificial Intellingence: Rules based AI and Learning based AI.

    Rule Based AI

    Imagine one of the rovers on Mars. What does it happen when an obstacol is in front of him? Simple, developers provided a set of rules to apply when something new happen: if you find an obstacol go back, move to the right and go forward. That’s exactly the same rule that our home cleaning robot follows in our houses.

    The most important thing is that not the rover on Mars, not our clieaning robot have been informed about the planimetry of the house or of Mars. So they have to improvise and learn where the obstacole is so to not run into it again.

    With Rule based AI we provide rules to the AI model, the model adapt to a new (never seen) input appling that rules and learn something new to be used for the future.

    This is the basis for games’ AI model such as chess, or Go.

    Learnig Based AI (a.k.a Machine Learning)

    What if I provide to an AI system the rule to… create yourself the rules extracting them from a very very very very large amount of data? Well, this is a Machine Learning algorithm (in a very very very very simple way).

    Let the machine find the connections among information provided so to be ready to categorize in the right way next (never seen) input.

    As you can imagine, this implies that training data need to be very very very very accurate and that you need also a set of testing data to verify the inferenced rules are… well, the right ones. This is called unsupervised learning.

    Maybe, you can also try to supervise the learning process so to correct some non valid inferenced rule. In that case is called superviced learning.

    When the tasks became really difficult that machine learning need to jump to Deep Learning, usign different algorithms. To semplify, if with a machine learning algorithm your AI model is able to indentify with a picture if there is a human or a plant, with deep learning algorithms your AI model is able to understand in the same picture the emotions that the human seems to feel. This is a complete new the level of complexity.

    Generative AI

    When a Deep Learning AI Model is used to create content, you have a Generative AI Model! Very simple.

    LLM (Large Language Model) is one of the Generative AI models.

    So, Generative AI models suffer from the same issues as Machine Learning and Deep Learning models. Bad training data? Then you will have lot of hallucinations (yes, this is the technical term to identify errors in contente generated by AI).

    And that’s it. In a very very very simple way!