part2-1.md

Part 2: Create the Retriever, Add Semantic Kernel and Plugins

In the second part we will create the RAG application. Fist let's work on the retrieval.

Create the DbRetriever

This section we will implement code using Entity Framework Core to perform a similarity search using the embeddings we populated in the last section.

1. In the PdfChatApp project folder, create a new folder named Retrievers and add a file named DbRetriever.cs to it.

This will hold the logic to do a similarity search in the database.

2. Paste the following code into the DbRetriever.cs file you just created:

using Microsoft.EntityFrameworkCore;
using PdfChatApp.Extensions;
using PdfChatApp.Models;
using SmartComponents.LocalEmbeddings.SemanticKernel;

namespace PdfChatApp.Retrievers;

public record DocumentInfo(int Id, int PageId, int ChunkId, string Name, string Path, string Text, double Similarity);

public class DbRetriever(DocDbContext dbContext)
{
    public async Task<List<DocumentInfo>> RetrieveLocalAsync(string text, int k)
    {
        using var localEmbeddingGenerator = new LocalTextEmbeddingGenerationService();

        var vector = (await localEmbeddingGenerator.GenerateEmbeddingsAsync([text])).ToFloatArray();

        return await dbContext.PageChunks
            .Include(pc => pc.Page).ThenInclude(p => p!.Document)
            .OrderBy(pc => EF.Functions.VectorDistance("cosine", pc.Embedding ?? new float[0], vector))
            .Take(k)
            .Select(pc => new DocumentInfo(pc!.Page!.Document!.Id, pc.PageId, pc.Id, pc.Page.Document.Name, pc.Page.Document.Path, pc.Text, EF.Functions.VectorDistance("cosine", pc.Embedding ?? new float[0], vector)))
            .ToListAsync();
    }
}

In the code, we first declare a DocumentInfo record to use in returning the information about the page chunk, document, text and similarity score.

The DbRetriever class has one method RetrieveLocalAsync() which uses the LocalTextEmbeddingGenerationService to get the embeddings for the passed in text, then uses the EF Core plugin mentioned earlier to use the EF.Functions.VectorDistance() to order the search results by their similarity score and returns the top k number of items.

Now we need to add Azure OpenAI and Semantic Kernel.

Create an Azure OpenAI Resource and a Deployment

In this section you will create an Azure OpenAI resource and deployment for using in the next section.

1. In the Azure Portal, type OpenAI in the search box in the top center

2. Select Azure OpenAI in the list.
3. Click the Create button in the upper left.

4. Fill out the following items:

• Subscription - select the subscription you want to use
• Resource Group - select the resource group you used with the Azure SQL resource
• Region - select the region you chose for the Azure SQL resource
• Name - provide a unique name for your Azure OpenAI resource
• Pricing tier - select the Standard S0 pricing tier

5. Click the Next button
6. On the Network tab, click Next

7. On the Tags tab, click Next
8. Click Create

This will create the Azure OpenAI resource. Next we need to create a gpt4o deployment.

Create a GPT4o deployment

1. Once your Azure OpenAI resource is created, click on the Go to resource button

2. In the upper left corner, click the Go to Azure OpenAI Studio button

This will open the Azure OpenAI Studio in a new tab. You may need to authenticate again.

3. Click the Deployments menu item on the lower left

4. Click the Deploy Model dropdown button and the Deploy base model option

5. On the Select a model dialog, find gpt-4o in the listing and select it.

6. Select the Confirm button on the dialog.
7. On the Deploy model dialog, click the Deploy button

This could take a few minutes, so while that is creating, lets start adding Semantic Kernel to the application.

Next: Add Semantic Kernel >