Semantic Kernel is Microsoft's open-source SDK for integrating language models into applications. Its central concept is the kernel: a container that brings together the AI service, the plugins with the application's functions and the configuration.

Creating the kernel

var builder = Kernel.CreateBuilder();
builder.AddOpenAIChatCompletion(
    modelId: "gpt-4o-mini",
    apiKey: configuration["OpenAI:ApiKey"]!);
builder.Plugins.AddFromType<WarehousePlugin>();

Kernel kernel = builder.Build();

A plugin

A plugin is a class with methods decorated with [KernelFunction]. Descriptions are essential: by reading them the model decides when to call a function.

public class WarehousePlugin(IWarehouse warehouse)
{
    [KernelFunction, Description("Returns the available quantity of an item given its code")]
    public Task<int> Stock([Description("Item code, e.g. ABC-123")] string code) =>
        warehouse.QuantityAsync(code);
}

Automatic function calling

var settings = new OpenAIPromptExecutionSettings
{
    FunctionChoiceBehavior = FunctionChoiceBehavior.Auto()
};

var answer = await kernel.InvokePromptAsync(
    "Do we have enough units of item ABC-123 for an order of 40?",
    new KernelArguments(settings));

Console.WriteLine(answer);

The model understands that it needs the stock level, calls Stock("ABC-123"), receives the number and formulates the answer. The application's code remains the one deciding and controlling the data.

Good practices

• Expose few functions, with clear names and descriptions: too many functions confuse the model.
• Always validate arguments and enforce the user's permissions inside the functions.
• For operations that modify data, ask the user for confirmation before executing them.
• Record traces and costs of calls: Semantic Kernel integrates with OpenTelemetry.

Semantic Kernel coexists with the Microsoft.Extensions.AI abstractions, and its evolution towards agents continues in the Microsoft Agent Framework.