Skip to main content

Kernel

The Kernel is the central orchestrator in Semantic Kernel that manages services, plugins, and execution flow.

Overview​

The Kernel acts as the main entry point and coordinator for AI operations in Semantic Kernel applications.

The kernel:

  • Serves as the core of the Semantic Kernel SDK.

  • Manages all services (AI services, logging, HTTP clients) and plugins essential for your application.

  • Provides a unified entry point—whether running prompts or custom code, the kernel is always available to retrieve needed components.

This centralization simplifies configuration, observability, and monitoring of AI agents and their workflows.

Lifecycle: How the Kernel Processes a Prompt​

When you invoke a prompt via the kernel, it orchestrates the following sequence:

  • Selects the most appropriate AI service.

  • Builds the prompt using a template.

  • Sends the prompt to the specified AI service.

  • Receives and parses the AI’s response.

  • Returns the response to your application.

Through this flow, you can attach events or middleware—for logging, status tracking, or responsible AI enforcement—at any stage.

Building the Kernel: Services & Plugins​

Components​

Services​

Include AI interfaces (like chat completion) or utility services (e.g., logging, HTTP clients). Modeled after the .NET Service Provider pattern for cross-language support. Microsoft Learn

Plugins​

Encapsulate work functions—e.g., data retrieval or API operations—that AI services and prompt templates may invoke.

Key Concepts​

1. Service Registration​

The Kernel manages various services through a service selector:

// Registering a HuggingFace TGI service
const tgi = HuggingFaceTGI.Builder()
  .withModelId("qwen3-1-7b-wrh")
  .withClient(
    new HuggingFaceClient({
      baseURL: "https://somewhere.us-east-1.aws.endpoints.huggingface.cloud/v1/",
      apiKey: "hf_XXXX", // inference token
    })
  )
  .build()

const kernel = Kernel.Builder().withAIService(HuggingFaceTGI, tgi).build()

2. Plugin Management​

Plugins extend the Kernel's capabilities with additional functions:

const myPlugin = KernelPluginFactory.createFromObject(new MyPlugin(), "MyPlugin")

const kernel = Kernel.Builder().withPlugin(conversationSummaryPlugin).build()

3. Function Invocation​

The Kernel enables execution of both native and semantic functions:

const myFunction = KernelFunctionFromPrompt.Builder()
  .withTemplate("My Function prompt")
  .withName("myFunction")
  .withExecutionSettings(
    PromptExecutionSettings.Builder<GeminiChatCompletionPromptExecutionSettings>()
      .maxOutputTokens(1000)
      .temperature(0.4)
      .topP(1)
      .build()
  )
  .build()

const result = await kernel.invoke(
  myFunction,
  KernelArguments.Builder().withInput("My Function").build(),
  InvocationContext.Builder().withServiceClass(GeminiChatCompletion).build()
)