175 lines
6.9 KiB
C#
175 lines
6.9 KiB
C#
// Copyright (c) Microsoft. All rights reserved.
|
|
|
|
using MCPServer;
|
|
using MCPServer.ProjectResources;
|
|
using MCPServer.Prompts;
|
|
using MCPServer.Resources;
|
|
using MCPServer.Tools;
|
|
using Microsoft.Extensions.AI;
|
|
using Microsoft.Extensions.VectorData;
|
|
using Microsoft.SemanticKernel;
|
|
using Microsoft.SemanticKernel.Agents;
|
|
using Microsoft.SemanticKernel.Connectors.InMemory;
|
|
using ModelContextProtocol.Protocol;
|
|
using ModelContextProtocol.Server;
|
|
|
|
var builder = Host.CreateEmptyApplicationBuilder(settings: null);
|
|
|
|
// Load and validate configuration
|
|
(string embeddingModelId, string chatModelId, string apiKey) = GetConfiguration();
|
|
|
|
// Register the kernel
|
|
IKernelBuilder kernelBuilder = builder.Services.AddKernel();
|
|
|
|
// Register SK plugins
|
|
kernelBuilder.Plugins.AddFromType<DateTimeUtils>();
|
|
kernelBuilder.Plugins.AddFromType<WeatherUtils>();
|
|
kernelBuilder.Plugins.AddFromType<MailboxUtils>();
|
|
|
|
// Register SK agent as plugin
|
|
kernelBuilder.Plugins.AddFromFunctions("Agents", [AgentKernelFunctionFactory.CreateFromAgent(CreateSalesAssistantAgent(chatModelId, apiKey))]);
|
|
|
|
// Register embedding generation service and in-memory vector store
|
|
kernelBuilder.Services.AddSingleton<VectorStore, InMemoryVectorStore>();
|
|
kernelBuilder.Services.AddOpenAIEmbeddingGenerator(embeddingModelId, apiKey);
|
|
|
|
// Register MCP server
|
|
builder.Services
|
|
.AddMcpServer()
|
|
.WithStdioServerTransport()
|
|
|
|
// Add all functions from the kernel plugins to the MCP server as tools
|
|
.WithTools()
|
|
|
|
// Register the `getCurrentWeatherForCity` prompt
|
|
.WithPrompt(PromptDefinition.Create(EmbeddedResource.ReadAsString("getCurrentWeatherForCity.json")))
|
|
|
|
// Register vector search as MCP resource template
|
|
.WithResourceTemplate(CreateVectorStoreSearchResourceTemplate())
|
|
|
|
// Register the cat image as a MCP resource
|
|
.WithResource(ResourceDefinition.CreateBlobResource(
|
|
uri: "image://cat.jpg",
|
|
name: "cat-image",
|
|
content: EmbeddedResource.ReadAsBytes("cat.jpg"),
|
|
mimeType: "image/jpeg"));
|
|
|
|
await builder.Build().RunAsync();
|
|
|
|
/// <summary>
|
|
/// Gets configuration.
|
|
/// </summary>
|
|
static (string EmbeddingModelId, string ChatModelId, string ApiKey) GetConfiguration()
|
|
{
|
|
// Load and validate configuration
|
|
IConfigurationRoot config = new ConfigurationBuilder()
|
|
.AddUserSecrets<Program>()
|
|
.AddEnvironmentVariables()
|
|
.Build();
|
|
|
|
if (config["OpenAI:ApiKey"] is not { } apiKey)
|
|
{
|
|
const string Message = "Please provide a valid OpenAI:ApiKey to run this sample. See the associated README.md for more details.";
|
|
Console.Error.WriteLine(Message);
|
|
throw new InvalidOperationException(Message);
|
|
}
|
|
|
|
string embeddingModelId = config["OpenAI:EmbeddingModelId"] ?? "text-embedding-3-small";
|
|
|
|
string chatModelId = config["OpenAI:ChatModelId"] ?? "gpt-4o-mini";
|
|
|
|
return (embeddingModelId, chatModelId, apiKey);
|
|
}
|
|
static ResourceTemplateDefinition CreateVectorStoreSearchResourceTemplate(Kernel? kernel = null)
|
|
{
|
|
return new ResourceTemplateDefinition
|
|
{
|
|
Kernel = kernel,
|
|
ResourceTemplate = new()
|
|
{
|
|
UriTemplate = "vectorStore://{collection}/{prompt}",
|
|
Name = "Vector Store Record Retrieval",
|
|
Description = "Retrieves relevant records from the vector store based on the provided prompt."
|
|
},
|
|
Handler = async (
|
|
RequestContext<ReadResourceRequestParams> context,
|
|
string collection,
|
|
string prompt,
|
|
[FromKernelServices] IEmbeddingGenerator<string, Embedding<float>> embeddingGenerator,
|
|
[FromKernelServices] VectorStore vectorStore,
|
|
CancellationToken cancellationToken) =>
|
|
{
|
|
// Get the vector store collection
|
|
VectorStoreCollection<Guid, TextDataModel> vsCollection = vectorStore.GetCollection<Guid, TextDataModel>(collection);
|
|
|
|
// Check if the collection exists, if not create and populate it
|
|
if (!await vsCollection.CollectionExistsAsync(cancellationToken))
|
|
{
|
|
static TextDataModel CreateRecord(string text, ReadOnlyMemory<float> embedding)
|
|
{
|
|
return new()
|
|
{
|
|
Key = Guid.NewGuid(),
|
|
Text = text,
|
|
Embedding = embedding
|
|
};
|
|
}
|
|
|
|
string content = EmbeddedResource.ReadAsString("semantic-kernel-info.txt");
|
|
|
|
// Create a collection from the lines in the file
|
|
await vectorStore.CreateCollectionFromListAsync<Guid, TextDataModel>(collection, content.Split('\n'), embeddingGenerator, CreateRecord);
|
|
}
|
|
|
|
// Generate embedding for the prompt
|
|
ReadOnlyMemory<float> promptEmbedding = (await embeddingGenerator.GenerateAsync(prompt, cancellationToken: cancellationToken)).Vector;
|
|
|
|
// Retrieve top three matching records from the vector store
|
|
var result = vsCollection.SearchAsync(promptEmbedding, top: 3, cancellationToken: cancellationToken);
|
|
|
|
// Return the records as resource contents
|
|
List<ResourceContents> contents = [];
|
|
|
|
await foreach (var record in result)
|
|
{
|
|
contents.Add(new TextResourceContents()
|
|
{
|
|
Text = record.Record.Text,
|
|
Uri = context.Params!.Uri!,
|
|
MimeType = "text/plain",
|
|
});
|
|
}
|
|
|
|
return new ReadResourceResult { Contents = contents };
|
|
}
|
|
};
|
|
}
|
|
|
|
static Agent CreateSalesAssistantAgent(string chatModelId, string apiKey)
|
|
{
|
|
IKernelBuilder kernelBuilder = Kernel.CreateBuilder();
|
|
|
|
// Register the SK plugin for the agent to use
|
|
kernelBuilder.Plugins.AddFromType<OrderProcessingUtils>();
|
|
|
|
// Register chat completion service
|
|
kernelBuilder.Services.AddOpenAIChatCompletion(chatModelId, apiKey);
|
|
|
|
// Using a dedicated kernel with the `OrderProcessingUtils` plugin instead of the global kernel has a few advantages:
|
|
// - The agent has access to only relevant plugins, leading to better decision-making regarding which plugin to use.
|
|
// Fewer plugins mean less ambiguity in selecting the most appropriate one for a given task.
|
|
// - The plugin is isolated from other plugins exposed by the MCP server. As a result the client's Agent/AI model does
|
|
// not have access to irrelevant plugins.
|
|
Kernel kernel = kernelBuilder.Build();
|
|
|
|
// Define the agent
|
|
return new ChatCompletionAgent()
|
|
{
|
|
Name = "SalesAssistant",
|
|
Instructions = "You are a sales assistant. Place orders for items the user requests and handle refunds.",
|
|
Description = "Agent to invoke to place orders for items the user requests and handle refunds.",
|
|
Kernel = kernel,
|
|
Arguments = new KernelArguments(new PromptExecutionSettings() { FunctionChoiceBehavior = FunctionChoiceBehavior.Auto() }),
|
|
};
|
|
}
|