Uncovering GenAI - Chapter 5: Building a Retrieval-Augmented Generation Application

            sequenceDiagram
                actor User
                participant App as Console
Application
(.NET)
                participant Data Source as Hugging Face
Star Wars
Characters Dataset
                participant Qdrant as Qdrant
Vector
Database
                participant Embed as Ollama
Embedding
Model
                participant LLM as Ollama
Chat
Model

                Note left of App: Phase 1 -
Indexing: Building the index
                App->>Data Source: Fetch Star
Wars Characters
Dataset
                App->>Embed: Convert prompt to
embedding
                Embed-->>App: Returns embedding
                App->>Qdrant: Upload embeddings
to vector database
                Qdrant-->>App: Confirms upload

                Note left of App: Phase 2 -
Retrieval i.e.
Responding to user prompt/input
                loop User Interaction
                    User->>App: Inputs a prompt/question
                    App->>Embed: Convert question to
embedding
                    Embed-->>App: Returns question
embedding
                    App->>Qdrant: Query vector database
for relevant data
                    Qdrant-->>App: Returns relevant data
                    App->>App: Create context
from question, data and
additional instructions
                    App->>LLM: Send context to
LLM for response
                    LLM-->>App: Returns response
                    App->>User: Display response
                end
                loop
                    User->>App: Inputs a prompt/question
                    App->>Embed: Convert question to
embedding
                    Embed-->>App: Returns question
embedding
                    App->>Qdrant: Query vector database
for relevant data
                    Qdrant-->>App: Returns relevant data
                    App->>App: Create context
from question, data and
additional instructions
                    App->>LLM: Send context to
LLM for response
                    LLM-->>App: Returns response
                    App->>User: Display response
                end
        

• Create a RagSemanticKernelChatApplication class and use the constructor to set up the application host and configurations. The configuration setup will allow reading various settings from appsettings.json in the next step.
• Now we will register a service with dependency injection that lets our application use a local Ollama embeddings and chat completion model (our downloaded LLM model in the Prerequisites step). This is done by leveraging the out-of-the-box methods from Semantic Kernel AddOllamaEmbeddingGenerator, AddQdrantVectorStore, and AddOllamaChatCompletion method that the Semantic Kernel packages.
• We will now register the Kernel class as a transient service in the dependency injection container. This will allow dependency injection to resolve the Kernel class on demand.
• Finally, we will build the host and expose a method to get the service from the host's service provider.

public class RagSemanticKernelChatApplication
{
    private readonly IHost _host;

    public RagSemanticKernelChatApplication()
    {
        var hostBuilder = Host.CreateDefaultBuilder();
        hostBuilder.ConfigureAppConfiguration(cfg => { cfg.AddJsonFile("appsettings.json", false, true); });

        hostBuilder.ConfigureServices((context, services) =>
        {
            services.AddSingleton(context.Configuration);
            RegisterDataSourceService(context.Configuration, services);
            RegisterEmbeddingGenerator(context.Configuration, services);
            RegisterVectorDb(context.Configuration, services);
            RegisterChatProvider(context.Configuration, services);
            RegisterKernel(services);
        });

        _host = hostBuilder.Build();
    }

    private static void RegisterDataSourceService(IConfiguration config, IServiceCollection services)
    {
        var starWarBaseUrl = new Uri(config["DataSource:Endpoint"] ?? string.Empty);
        services.AddHttpClient<IStarWarCharacterService, StarWarCharacterService>(c =>
            c.BaseAddress = starWarBaseUrl);
    }

    private static void RegisterEmbeddingGenerator(IConfiguration config, IServiceCollection services)
    {
        var embeddingEndpoint = new Uri(config["Ollama:Endpoint"] ?? string.Empty);
        var model = config["Ollama:EmbeddingModel"] ?? string.Empty;
        services.AddOllamaEmbeddingGenerator(model, embeddingEndpoint);
    }

    private void RegisterVectorDb(IConfiguration config, IServiceCollection services)
    {
        services.AddQdrantVectorStore(
            config["Qdrant:Endpoint"] ?? "",
            int.Parse(config["Qdrant:Port"] ?? "6334"),
            false
        );
    }

    private void RegisterChatProvider(IConfiguration config, IServiceCollection services)
    {
        var chatEndpoint = new Uri(config["Ollama:Endpoint"] ?? string.Empty);
        var model = config["Ollama:ChatModel"] ?? string.Empty;
        services.AddOllamaChatCompletion(
            model,
            chatEndpoint
        );
    }

    private void RegisterKernel(IServiceCollection services)
    {
        services.AddTransient(serviceProvider => new Kernel(serviceProvider));
    }

    public T GetService<T>() where T : notnull
    {
        return _host.Services.GetRequiredService<T>();
    }
}
            

Console.ForegroundColor = ConsoleColor.Green;
Console.WriteLine("Hello, I am a Star Wars character expert. How can I help you today?");
Console.ResetColor();

var app = new RagSemanticKernelChatApplication();

var characterService = app.GetService<IStarWarCharacterService>();
var characters = await characterService.GetCharactersAsync();


var kernel = app.GetService<Kernel>();
var assistant = new CharacterAssistant(kernel);
await assistant.IngestAsync(characters);

do
{
    Console.WriteLine("==========================================================");
    Console.WriteLine("User: ");
    var userInput = Console.ReadLine();
    if (string.IsNullOrWhiteSpace(userInput) || userInput.Equals("exit", StringComparison.OrdinalIgnoreCase))
    {
        Console.WriteLine("Exiting chat...");
        break;
    }

    var reply = await assistant.ReplyToPrompt(userInput);
    Console.ForegroundColor = ConsoleColor.Green;
    Console.WriteLine("AI:");
    foreach (var line in reply.Split("\n")) Console.WriteLine($"   {line}");
    Console.ResetColor();
} while (true);