To complete this chapter, you will need to have a recent, supported version of your preferred Python or C# development environment:

• For Python, the minimum supported version is Python 3.10, and the recommended version is Python 3.11
• For C#, the minimum supported version is .NET 8

Important note

The examples are presented in C# and Python, and you can choose to only read the examples of your preferred language. Occasionally, a feature is available in only one of the languages. In such cases, we provide an alternative in the other language for how to achieve the same objectives.

In this chapter, we will call OpenAI services. Given the amount that companies spend on training these large language models (LLMs), it’s no surprise that using these services is not free. You will need an OpenAI API key, obtained either directly through OpenAI or Microsoft, via the Azure OpenAI service.

Important: Using the OpenAI services is not free...

Generative AI refers to a subset of artificial intelligence programs that are capable of creating content that is similar to what humans can produce. These systems use training from very large datasets to learn their patterns, styles, and structures. Then, they can generate entirely new content, such as synthesized images, music, and text.

Using GenAI as consumer or end-user is very easy, and as a technical person, you probably have already done it. There are many consumer-facing AI products. The most famous is OpenAI’s ChatGPT, but there are many others that have hundreds of millions of users every day, such as Microsoft Copilot, Google Gemini (formerly Bard), and Midjourney. As of October 2023, Meta, the parent company of Facebook, WhatsApp, and Instagram, is making GenAI services available to all its users, increasing the number of GenAI daily users to billions.

While the concept of GenAI has existed for a while, it gained a lot of users...

Microsoft Semantic Kernel (https://github.com/microsoft/semantic-kernel) is a thin, open source software development toolkit (SDK) that makes it easier for applications developed in C# and Python to interact with AI services such as the ones made available through OpenAI, Azure OpenAI, and Hugging Face. Semantic Kernel can receive requests from your application and route them to different AI services. Furthermore, if you extend the functionality of Semantic Kernel by adding your own functions, which we will explore in Chapter 3, Semantic Kernel can automatically discover which functions need to be used, and in which order, to fulfill a request. The request can come directly from the user and be passed through directly from your application, or your application can modify and enrich the user request before sending it to Semantic Kernel.

It was originally designed to power different versions of Microsoft Copilot, such as Microsoft 365 Copilot and the Bing...

To complete this section, you must have an API key. The process to obtain an API key was described at the beginning of this chapter.

In the upcoming subsections, we are only going to connect to the OpenAI text models GPT-3.5 and GPT-4. If you have access to the OpenAI models through Azure, you will need to make minor modifications to your code.

Although it would be simpler to connect to a single model, we are already going to show a simple but powerful Microsoft Semantic Kernel feature: we’re going to connect to two different models and run a simple prompt using the simpler but less expensive model, GPT-3.5, and a more complex prompt on the more advanced but also more expensive model, GPT-4.

This process of sending simpler requests to simpler models and more complex requests to more complex models is something that you will frequently do when creating your own applications. This approach is called LLM cascade, and it was...

This section assumes you completed the prior sections and builds upon the same code. By now, you should have instantiated Semantic Kernel and loaded both the GPT-3.5 and the GPT-4 services into it in that order. When you submit a prompt, it will default to the first service, and will run the prompt on GPT-3.5.

When we send the prompt to the service, we will also send a parameter called temperature. The temperature parameter goes from 0.0 to 1.0, and it controls how random the responses are. We’re going to explain the temperature parameter in more detail in later chapters. A temperature parameter of 0.8 generates a more creative response, while a temperature parameter of 0.2 generates a more precise response.

To send the prompt to the service, we will use a method called create_semantic_function. For now, don’t worry about what a semantic function is. We’re going to explain it in the Using generative AI to solve simple problems section...

Microsoft Semantic Kernel distinguishes between two types of functions that can be loaded into it: semantic functions and native functions.

Semantic functions are functions that connect to AI services, usually LLMs, to perform a task. The service is not part of your codebase. Native functions are regular functions written in the language of your application.

The reason to differentiate a native function from any other regular function in your code is that the native function will have additional attributes that will tell the kernel what it does. When you load a native function into the kernel, you can use it in chains that combine native and semantic functions. In addition, Semantic Kernel planner can use the function when creating a plan to achieve a user goal.

Creating semantic functions

We have already created a semantic function (knock) in the previous section. Now, we’re going to add a parameter to it. The default...

One of the greatest strengths of Microsoft Semantic Kernel is that you can create semantic plugins that are language agnostic. Semantic plugins are collections of semantic functions that can be imported into the kernel. Creating semantic plugins allows you to separate your code from the AI function, which makes your application easier to maintain. It also allows other people to work on the prompts, making it easier to implement prompt engineering, which will be explored in Chapter 2.

Each function is defined by a directory containing two text files: config.json, which contains the configuration for the semantic function, and skprompt.txt, which contains its prompt.

The configuration of the semantic function includes the preferred engine to use, the temperature parameter, and a description of what the semantic function does and its inputs.

The text file contains the prompt that will be sent to the AI service to generate the response.

In this section, we are going...

Instead of calling functions yourself, you can let Microsoft Semantic Kernel choose the functions for you. This can make your code a lot simpler and can give your users the ability to combine your code in ways that you haven’t considered.

Right now, this will not seem very useful because we only have a few functions and plugins. However, in a large application, such as Microsoft Office, you may have hundreds or even thousands of plugins, and your users may want to combine them in ways that you can’t yet imagine. For example, you may be creating a copilot that helps a user be more efficient when learning about a subject, so you write a function that downloads the latest news about that subject from the web. You may also have independently created a function that explains a piece of text to the user so that the user can paste content to learn more about it. The user may decide to combine them both with “download the news...

In this chapter, you learned about Generative AI and the main components of Microsoft Semantic Kernel. You learned how to create a prompt and submit it to a service and how to embed that prompt into a semantic function. You also learned how to execute multistep requests by using a planner.

In the next chapter, we are going to learn how to make our prompts better through a topic called prompt engineering. This will help you create prompts that get your users the correct result faster and use fewer tokens, therefore reducing costs.

[1] A. Vaswani et al., “Attention Is All You Need,” Jun. 2017.

[2] OpenAI, “GPT-4 Technical Report.” arXiv, Mar. 27, 2023. doi: 10.48550/arXiv.2303.08774.

[3] L. Chen, M. Zaharia, and J. Zou, “FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance.” arXiv, May 09, 2023. doi: 10.48550/arXiv.2305.05176.