Chatbot:
  User: "What are my pending orders?"
  Bot:  "I don't have access to your orders. Please check your account page."

Agent:
  User: "What are my pending orders?"
  Agent → calls get_orders(userId) tool
  Agent → reads result: [Order #1234, Order #5678]
  Agent: "You have 2 pending orders: #1234 (ships tomorrow) and #5678 (processing)."

dotnet add package Microsoft.SemanticKernel

// Setup
var kernel = Kernel.CreateBuilder()
    .AddOpenAIChatCompletion(
        modelId:   "gpt-4o",
        apiKey:    Environment.GetEnvironmentVariable("OPENAI_API_KEY")!)
    .Build();

// Define a plugin — functions the AI can call
public class OrderPlugin(IOrderRepository orders)
{
    [KernelFunction("get_pending_orders")]
    [Description("Gets a list of pending orders for a user")]
    public async Task<string> GetPendingOrdersAsync(
        [Description("The user ID")] string userId)
    {
        var pending = await orders.GetPendingAsync(userId);
        return JsonSerializer.Serialize(pending);
    }

    [KernelFunction("cancel_order")]
    [Description("Cancels an order by ID. Returns success or error message.")]
    public async Task<string> CancelOrderAsync(
        [Description("The order ID to cancel")] string orderId)
    {
        var result = await orders.CancelAsync(orderId);
        return result ? $"Order {orderId} cancelled successfully" : $"Failed to cancel {orderId}";
    }
}

// Register the plugin
kernel.Plugins.AddFromObject(new OrderPlugin(orderRepo), "Orders");

// Chat with auto-invocation of tools
var chat = kernel.GetRequiredService<IChatCompletionService>();

var history = new ChatHistory();
history.AddSystemMessage(
    "You are a helpful order management assistant. " +
    "Use the available tools to answer questions about orders.");

history.AddUserMessage("Show me my pending orders for user U-001");

var settings = new OpenAIPromptExecutionSettings
{
    ToolCallBehavior = ToolCallBehavior.AutoInvokeKernelFunctions  // agent mode
};

var reply = await chat.GetChatMessageContentAsync(history, settings, kernel);
Console.WriteLine(reply.Content);
// Output: "You have 2 pending orders: ..."

// Add vector memory for searching your own data
var memory = new SemanticTextMemory(
    new AzureAISearchMemoryStore(endpoint, apiKey),
    new AzureOpenAITextEmbeddingGenerationService(deploymentName, endpoint, apiKey)
);

// Store documents
await memory.SaveInformationAsync(
    collection: "docs",
    id:         "policy-001",
    text:       "Returns are accepted within 30 days of purchase.",
    description: "Return policy"
);

// The model can query memory via a plugin
kernel.ImportPluginFromObject(
    new TextMemoryPlugin(memory), "memory");

// Handlebars planner — AI generates a plan as Handlebars template
var planner    = new HandlebarsPlanner();
var goal       = "Summarize the top 3 pending orders for user U-001 and email a report to manager@company.com";
var plan       = await planner.CreatePlanAsync(kernel, goal);

Console.WriteLine(plan);  // shows the generated plan

var result = await plan.InvokeAsync(kernel);
Console.WriteLine(result);

// Define specialized agents
var orderAgent = new ChatCompletionAgent
{
    Kernel       = kernel,
    Name         = "OrderAgent",
    Instructions = "You manage orders. Use the Orders plugin.",
};

var emailAgent = new ChatCompletionAgent
{
    Kernel       = kernel,
    Name         = "EmailAgent",
    Instructions = "You send emails. Use the Email plugin.",
};

// Orchestrate with AgentGroupChat
var chat = new AgentGroupChat(orderAgent, emailAgent)
{
    ExecutionSettings = new AgentGroupChatSettings
    {
        TerminationStrategy = new ApprovalTerminationStrategy()
    }
};

await foreach (var message in chat.InvokeAsync())
    Console.WriteLine($"{message.AuthorName}: {message.Content}");

RAG (Retrieval-Augmented Generation):
  • Query a vector store, inject results into the prompt
  • Best for: Q&A over static documents, knowledge bases

Agent:
  • LLM calls tools dynamically based on the task
  • Best for: multi-step tasks, real-time data, taking actions

Fine-tuning:
  • Train the model on your domain data
  • Best for: specific tone/format, specialized knowledge, cheaper inference
  • Expensive and rarely necessary