We’ll continue to implement the architecture of GenAISys as we’ve defined in Figure 1.1 from Chapter 1. Figure 2.1, on the other hand, takes us further into the underlying functions of a GenAISys.

Figure 2.1: Defining the functions to build

We established in the previous chapter that human roles are essential, and the preceding figure acknowledges that fact. We are the core of a GenAISys, no matter how advanced the building blocks (models or frameworks) are. Our first task is designing using our human creativity to find effective ways to implement a GenAISys controller. GenAISys needs human creativity, judgment, and technical decision-making. Under the hood of seamless copilots such as ChatGPT, Gemini, and Microsoft Copilot lie intricate layers of AI and non-AI logic. If we want to build our own ChatGPT-like system, we humans need to do the heavy lifting!

We will build two separate programs:

• A conversational...

Our two primary goals for this section are to build a conversational AI agent with the following:

• Short-term memory retention for a full ChatGPT-like conversational loop. The user and agent can have as many exchanges as they wish; there is no limit to the number of interactions between them.
• Long-term memory retention across multiple users and sessions. We’ll store in-memory sessions and persist them to a memory storage (in this case, a text file). This will enable multi-user contextual awareness for users such as John, Myriam, and Bob. Our conversational agent will move beyond classic one-to-one ChatGPT-style dialogues toward a custom GenAISys capable of handling multi-session, multi-user interactions.

To get started, open Conversational_AI_Agent.ipynb in this chapter’s GitHub directory (https://github.com/Denis2054/Building-Business-Ready-Generative-AI-Systems/tree/main). This notebook will guide you through the environment...

In this section, we will build the first component of our AI controller orchestrator: the ability to select the right task to perform. We develop this component as a standalone component that we will integrate starting from Chapter 3, where we will bridge the conversational agent with the AI controller orchestrator through a Pinecone vector store.

Figure 2.6 illustrates the workflow of the AI controller orchestrator we’ll be developing:

• C1. AI controller entry point input triggers the process.
• C2. Analyzes input, which could be a system or human user prompt.
• C3. Embeds user input through GPT-4o’s native functionality.
• C4. Embeds task scenario repository through GPT-4o’s native functionality.
• C5. Selects a scenario to execute a task that best matches the input.
• C6. Executes the scenario selected by the AI controller orchestrator.

Figure 2.6: Workflow of the AI controller orchestrator...

The first takeaway from this chapter is the central role of humans in a GenAISys. Human design drove the creation of both our conversational agent and orchestrator. We started developing these two complex components with simply an OpenAI API and Python, yet we humans designed the initial levels of the AI controller that powers our custom GenAISys. The basic GenAISys rule will always apply: no human roles, no GenAISys. We design AI systems, implement them, maintain them, and evolve them based on ongoing feedback.

The second takeaway is how our conversational AI agent goes beyond a small-scale ChatGPT-like structure. We not only built short-term context and memory retention for a full-turn dialogue, but we also added long-term memory across multiple users and multiple topics. Our dialogue included three users (John, Myriam, and Bob) and two topics (geology and surfing). As we progress through the book, we will expand the scope of these multi-user, multi-topic sessions to...

1. A ChatGPT-like GenAISys only needs a generative AI model such as GPT-4o. (True or False)
2. A ChatGPT-like GenAISys doesn’t require an AI controller. (True or False)
3. Human roles are critical when building and running GenAISys. (True or False)
4. Generally, not always, a generative AI model such as GPT-4o contains a task tag in one form or the other. (True or False)
5. Sometimes, not always, a generative model can find the most probable task to perform without a task tag. (True or False)
6. Semantic text similarity cannot be natively performed by GPT-4o. (True or False)
7. A full-turn generative AI conversation loop with an OpenAI API AI requires coding. (True or False)
8. Long-term memory AI conversation sessions are never necessary. (True or False)
9. Summarizing a text can only be done in English by GPT-4o. (True or False)
10. An AI controller orchestrator is sentient. (True or False)

• Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2020). Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. https://arxiv.org/abs/1910.10683
• Ren, J., Sun, Y., Du, H., Yuan, W., Wang, C., Wang, X., Zhou, Y., Zhu, Z., Wang, F., & Cui, S. (2024). Generative Semantic Communication: Architectures, Technologies, and Applications. https://doi.org/10.48550/arXiv.2412.08642

• Koziolek, H., Gruener, S., & Ashiwal, V. (2023). ChatGPT for PLC/DCS Control Logic Generation. https://doi.org/10.48550/arXiv.2305.15809

New frameworks, evolving architectures, research drops, production breakdowns—AI_Distilled filters the noise into a weekly briefing for engineers and researchers working hands-on with LLMs and GenAI systems. Subscribe now and receive a free eBook, along with weekly insights that help you stay focused and informed.

Subscribe at https://packt.link/TRO5B or scan the QR code below.