Basic

    %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart LR
        user(["User"]) -- input --> llm["LLM"]

        llm -- output --> user
        llm@{ shape: hex}
                            

var kernel = Kernel.CreateBuilder()
    .AddOllamaChatCompletion("gemma3:4b", new Uri("http://localhost:11434"))
    .Build();

var chat = kernel.GetRequiredService();

while (true)
{
    Console.Write("You: ");
    var question = Console.ReadLine();

    var result = await chat.GetChatMessageContentsAsync(question ?? "");

    Console.WriteLine("LLM: " + result.First().Content ?? "");
}
    

Chat history

              %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart LR
        user(["User"]) -- input --> history
        history["ChatHistory"] -- context --> llm["LLM"]

        llm -- add output --> history
        history -- output --> user
        history@{ shape: cyl}
        llm@{ shape: hex}
            

var kernel = Kernel.CreateBuilder()
    .AddOllamaChatCompletion("gemma3:4b", new Uri("http://localhost:11434"))
    .Build();

var chatCompletionService = kernel.GetRequiredService<IChatCompletionService>();
var chatHistory = new ChatHistory();
chatHistory.AddSystemMessage(
    """
    You are a helpful metal music expert.
    Your task is to provide the best music advice and facts about metal music for the user.
    Only metal music of course. Hell yeah! 🤘
    """
);
chatHistory.AddUserMessage([
    new TextContent("Which band invented metal? Just give the band name, no explanation.")
]);
var chatResult = await chatCompletionService.GetChatMessageContentAsync(chatHistory);
    

Templating (with prompty.ai)

              %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart LR
        user(["User"]) -- input --> template
        template["Template"] --> llm["LLM"]

        llm -- output --> user
        template@{ shape: doc}
        llm@{ shape: hex}
            

---
name: Metal_music_assistant
description: A prompt that leads to proper music advice 🎸
model:
  api: chat
  configuration:
    name: gemma3:4b
sample:
  question: "Which band invented metal?"
---

system:
You are a helpful metal music expert.
Your task is to provide the best music advice and facts about metal music for the user.
Only metal music of course. Hell yeah! 🤘
Never give an explanation unless explicitly asked.

user:
{{question}}
    

Functions (tools)

    %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart LR
        user(["User"]) -- input --> llm["LLM"]

        llm -- output --> user
        llm -- call --> tool
        llm@{ shape: hex}
                tool@{ shape: lin-rect}
            

        
public class MusicPlayerPlugin
{
    [KernelFunction("play_song")]
    public void PlaySong(string artist, string song)
    {
        Console.WriteLine($"MusicPlayerPlugin: Playing {song} by {artist}");
    }
}
    

var kernelBuilder = Kernel.CreateBuilder()
    .AddOllamaChatCompletion("llama3.1:8b", new Uri("http://localhost:11434"));

kernelBuilder.Plugins.AddFromType<MusicPlayerPlugin>("PlaySong");

var kernel = kernelBuilder.Build();
    

Model Context Protocol

• Tools
• Resources
• Prompts
• Sampling

Model Context Protocol

    %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart LR
        user(["User"]) -- input --> llm["LLM"]

        llm -- output --> user
        llm -- call --> mcpclient["MCP (tool) Client"]
        mcpclient -- http/stdio --> mcpserver["MCP Server"]
        llm@{ shape: hex}
            

var kernelBuilder = Kernel.CreateBuilder()
    .AddOllamaChatCompletion("llama3.1:8b", new Uri("http://localhost:11434"));

var transport = new SseClientTransport(new SseClientTransportOptions
{
    Endpoint = new Uri("http://localhost:3001"),
    UseStreamableHttp = true,
});

var mcpClient = await McpClientFactory.CreateAsync(transport);
var tools = await mcpClient.ListToolsAsync();

kernel.Plugins.AddFromFunctions("McpTools", tools.Select(f => f.AsKernelFunction()));

var kernel = kernelBuilder.Build();
    

RAG

Retrieval Augmented Generation

    %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart LR
        user(["User"]) -- input --> p1[/embed/]
        p1 -- input --> embedllm
        embedllm -- vector --> p1
        p1 -- vector --> db[(Store)]
        db -- best matches --> p2{{sfd}}
        user -- input --> p2[/create prompt/]
        p2 -- prompt -->llm
        llm -- output --> user
        llm["Query LLM"]
        embedllm["Embed LLM"]
        embedllm@{ shape: hex}
        llm@{ shape: hex}
            

3d vector

Vector search

Embeddings are vectors

Lorem ipsum dolor sit amet, consectetur adipiscing

Estimating story points, the data

• HuggingFace 🤗
• 20,5k rows
• Natural language user stories

Agent framework

Concurrent orchestration

    %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart TD
        i(["input"]) --> a1["Agent 1"]
        i --> a2["Agent 2"]
        i --> a3["Agent 3"]
        a1 --> c["Collector (aggregates)"]
        a2 --> c
        a3 --> c
        c --> o(["output"])
        a1@{ shape: hex}
        a2@{ shape: hex}
        a3@{ shape: hex}
            

Concurrent orchestration

Kernel kernel = ...;

ChatCompletionAgent physicist = new ChatCompletionAgent{
    Name = "PhysicsExpert",
    Instructions = "You are an expert in physics. You answer from physics perspective."
    Kernel = kernel,
};

ChatCompletionAgent chemist = new ChatCompletionAgent{
    Name = "ChemistryExpert",
    Instructions = "You are an expert in chemistry. You answer from chemistry perspective."
    Kernel = kernel,
};

ConcurrentOrchestration orchestration = new (physicist, chemist);

InProcessRuntime runtime = new InProcessRuntime();
await runtime.StartAsync();

var result = await orchestration.InvokeAsync("What is temperature?", runtime);
    

Sequential orchestration

                        %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
                        flowchart TD
                        i(["input"]) --> a1["Agent 1"]
                        a1 --> a2["Agent 2"]
                        a2 --> a3["Agent 3"]
                        a3 --> o(["output"])
                        a1@{ shape: hex}
                        a2@{ shape: hex}
                        a3@{ shape: hex}
            

Handoff orchestration

    %%{init: {'theme': 'dark', 'themeVariables': { 'darkMode': true }}}%%
    flowchart TD
        i(["input"]) --> a1["Agent 1"]
        h1["Human 1"] <--> a1
        a1 -- handoff --> a2["Agent 2"]
        a1 -- handoff --> a3["Agent 3"]
        a1 -- done --> o(["output"])
        a2 -- done --> o
        a3 -- done --> o
        a1@{ shape: hex}
        a2@{ shape: hex}
        a3@{ shape: hex}
            

Handoff orchestration

var handoffs = OrchestrationHandoffs
    .StartWith(triageAgent)
    .Add(triageAgent, statusAgent, returnAgent)
    .Add(statusAgent, triageAgent, "Transfer to this agent if the issue is not status related")
    .Add(returnAgent, triageAgent, "Transfer to this agent if the issue is not return related");

HandoffOrchestration orchestration = new HandoffOrchestration(
    handoffs,
    triageAgent,
    statusAgent,
    returnAgent)
{
    InteractiveCallback = interactiveCallback,
    ResponseCallback = responseCallback,
};

ValueTask interactiveCallback()
{
    var input = Console.ReadLine();
    return ValueTask.FromResult(new ChatMessageContent(AuthorRole.User, input));
}

InProcessRuntime runtime = new InProcessRuntime();
await runtime.StartAsync();

var result = await orchestration.InvokeAsync("I need help with my orders", runtime);
    

Magentic orchestration

Process framework

Do you really want that big of an (AI) framework dependency?