Semantic Kernel v2: Building Composable AI Applications in C#

Semantic Kernel v2 introduces a refined programming model for building AI applications in .NET. The new architecture emphasizes composability, testability, and enterprise patterns. Here’s how to leverage these improvements in your applications.

Setting Up the Kernel

The v2 API provides cleaner dependency injection and configuration:

using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Connectors.AzureOpenAI;

var builder = Kernel.CreateBuilder();

// Add AI services
builder.AddAzureOpenAIChatCompletion(
    deploymentName: "gpt-4o",
    endpoint: configuration["AzureOpenAI:Endpoint"],
    apiKey: configuration["AzureOpenAI:ApiKey"]
);

// Add plugins
builder.Plugins.AddFromType<CustomerPlugin>();
builder.Plugins.AddFromType<OrderPlugin>();

// Configure filters for observability
builder.Services.AddSingleton<IPromptRenderFilter, LoggingPromptFilter>();
builder.Services.AddSingleton<IFunctionInvocationFilter, MetricsFilter>();

var kernel = builder.Build();

Creating Plugins with Native Functions

Plugins encapsulate domain logic that the AI can invoke:

public class CustomerPlugin
{
    private readonly ICustomerService _customerService;

    public CustomerPlugin(ICustomerService customerService)
    {
        _customerService = customerService;
    }

    [KernelFunction("GetCustomerDetails")]
    [Description("Retrieves customer information by customer ID")]
    public async Task<CustomerInfo> GetCustomerAsync(
        [Description("The unique customer identifier")] string customerId)
    {
        return await _customerService.GetByIdAsync(customerId);
    }

    [KernelFunction("UpdateCustomerPreferences")]
    [Description("Updates customer communication preferences")]
    public async Task<bool> UpdatePreferencesAsync(
        string customerId,
        [Description("Email opt-in preference")] bool emailOptIn,
        [Description("SMS opt-in preference")] bool smsOptIn)
    {
        return await _customerService.UpdatePreferencesAsync(
            customerId, emailOptIn, smsOptIn);
    }
}

Orchestrating with Planners

The new Handlebars planner creates execution plans dynamically:

var planner = new HandlebarsPlanner(new HandlebarsPlannerOptions
{
    AllowLoops = true,
    MaxTokens = 4096
});

var plan = await planner.CreatePlanAsync(kernel,
    "Find customer John Smith and update their email preferences to opt-in");

var result = await plan.InvokeAsync(kernel);

Semantic Kernel v2 brings enterprise-grade patterns to AI development, making it easier to build maintainable, testable AI applications.