セマンティック カーネル C# 統合を使い始める

using Microsoft.Extensions.AI;
using Microsoft.Extensions.VectorData;
using Microsoft.SemanticKernel.Connectors.MongoDB;
using Microsoft.SemanticKernel.Data;
using MongoDB.Bson;
using MongoDB.Driver;
using OpenAI;
#pragma warning disable SKEXP0001
static class Program
{
    static async Task Main(string[] args)
    {
        // Get connection string and OpenAI API Key
        var connectionString = Environment.GetEnvironmentVariable("ATLAS_CONNECTION_STRING");
        if (connectionString == null)
        {
            Console.WriteLine("You must set your 'ATLAS_CONNECTION_STRING' environment variable.");
            Environment.Exit(0);
        }
        var openAIKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY");
        if (openAIKey == null)
        {
            Console.WriteLine("You must set your 'OPENAPI_KEY' environment variable.");
            Environment.Exit(0);
        }
        // Create new OpenAI API Embedding Model
        var embeddingGenerator = new OpenAIClient(openAIKey)
            .GetEmbeddingClient("text-embedding-ada-002")
            .AsIEmbeddingGenerator();
        // Instantiate Atlas as a vector store
        var mongoClient = new MongoClient(connectionString);
        var options = new MongoVectorStoreOptions { EmbeddingGenerator = embeddingGenerator };
        var vectorStore = new MongoVectorStore(mongoClient.GetDatabase("semantic_kernel_db"), options);
        // Sample data
        string[] lines =
        [
            "I am a developer",
            "I started using MongoDB two years ago",
            "I'm using MongoDB Vector Search with Semantic Kernel to implement RAG",
            "I like coffee"
        ];
        // Populate database with sample data
        await CreateCollectionFromListAsync<string, DataModel>(vectorStore, "records", lines, embeddingGenerator, CreateRecord);
        
        // Get the specific collection from the vector store
        var recordCollection = vectorStore.GetCollection<string, DataModel>("records");
    }
    static DataModel CreateRecord(string text, ReadOnlyMemory<float> embedding)
        => new()
        {
            Key = ObjectId.GenerateNewId().ToString(),
            Text = text,
            Embedding = embedding
        };
    static async Task CreateCollectionFromListAsync<TKey, TRecord>(
        this VectorStore vectorStore,
        string collectionName,
        string[] entries,
        IEmbeddingGenerator<string, Embedding<float>> embeddingGenerator,
        Func<string, ReadOnlyMemory<float>, TRecord> createRecord)
        where TKey : notnull
        where TRecord : class
    {
        // Get and create collection if it doesn't exist
        var collection = vectorStore.GetCollection<TKey, TRecord>(collectionName);
        await collection.EnsureCollectionExistsAsync().ConfigureAwait(false);
        // Create records and generate embeddings for them
        var embeddings = await embeddingGenerator.GenerateAsync(entries);
        var records = entries.Zip(embeddings, (entry, embedding) => createRecord(entry, embedding.Vector));
        // Add them to the database
        await collection.UpsertAsync(records).ConfigureAwait(false);
    }
    internal sealed class DataModel
    {
        [VectorStoreKey]
        [TextSearchResultName]
        public required String Key { get; init; }
        [VectorStoreData]
        [TextSearchResultValue]
        public required string Text { get; init; }
        [VectorStoreVector(1536)]
        public ReadOnlyMemory<float> Embedding { get; init; }
    }
}

dotnet run

1
// Create a text search instance using the InMemory vector store.
2
var textSearch = new VectorStoreTextSearch<DataModel>(recordCollection, embeddingGenerator);
3
4
// Search and return results as TextSearchResult items
5
var query = "What is my job title?";
6
KernelSearchResults<TextSearchResult> textResults = await textSearch.GetTextSearchResultsAsync(query, new() { Top = 2, Skip = 0 });
7
await foreach (TextSearchResult result in textResults.Results)
8
{
9
   Console.WriteLine($"Answer: {result.Value}");
10
}
11
Console.WriteLine("Search completed.");

// Create a kernel with OpenAI chat completion
IKernelBuilder kernelBuilder = Kernel.CreateBuilder();
kernelBuilder.AddOpenAIChatCompletion(
        modelId: "gpt-4o",
        apiKey: openAIKey);
Kernel kernel = kernelBuilder.Build();
// Create a text search instance using the vector store collection.
var textSearch = new VectorStoreTextSearch<DataModel>(recordCollection, embeddingGenerator);
// --- Modified RAG Section ---
var userQuestion = "When did I start using MongoDB?";
string retrievedContext = "No relevant context found."; // Default
// 1. Perform search to get context
var searchResults = await textSearch.GetTextSearchResultsAsync(userQuestion, new() { Top = 1 }); // Get most relevant result
await foreach (var result in searchResults.Results)
{
    if (result.Value != null)
    {
        retrievedContext = result.Value; // Use the text from the search result as context
        break; // Take the most relevant result
    }
}
// 2. Define a prompt template that uses the retrieved context
const string ragPromptTemplate = @"
Context:
{{$context}}
Question:
{{$question}}
Based *only* on the context provided, answer the question.
Answer:
";
// 3. Create a function from the RAG prompt template
var ragFunction = kernel.CreateFunctionFromPrompt(ragPromptTemplate);
// 4. Prepare arguments for the RAG prompt
var ragArguments = new KernelArguments
{
    ["question"] = userQuestion,
    ["context"] = retrievedContext
};
// 5. Invoke the RAG prompt
var ragResult = await kernel.InvokeAsync(ragFunction, ragArguments);
Console.WriteLine($"Question: {userQuestion}");
Console.WriteLine($"Retrieved Context: {retrievedContext}");
Console.WriteLine($"Answer: {ragResult.GetValue<string>()}");
// --- End of Modified RAG Section ---