Using a text embedding model locally with semantic kernel

Question

I've been reading Stephen Toub's blog post about building a simple console-based .NET chat application from the ground up with semantic-kernel. I'm following the examples but instead of OpenAI I want to use microsoft Phi 3 and the nomic embedding model. The first examples in the blog post I could recreate using the semantic kernel huggingface plugin. But I can't seem to run the text embedding example.

I've downloaded Phi and nomic embed text and are running them on a local server with lm studio.

Here's the code I came up with that uses the huggingface plugin:

using System.Net;
using System.Text;
using System.Text.RegularExpressions;
using Microsoft.Extensions.DependencyInjection;
using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Embeddings;
using Microsoft.SemanticKernel.Memory;
using System.Numerics.Tensors;
using Microsoft.Extensions.DependencyInjection;
using Microsoft.Extensions.Logging;
using Microsoft.SemanticKernel.ChatCompletion;

#pragma warning disable SKEXP0070, SKEXP0003, SKEXP0001, SKEXP0011, SKEXP0052, SKEXP0055, SKEXP0050  // Type is for evaluation purposes only and is subject to change or removal in future updates. 

internal class Program
{
    private static async Task Main(string[] args)
    {
        //Suppress this diagnostic to proceed.
        // Initialize the Semantic kernel
        IKernelBuilder kernelBuilder = Kernel.CreateBuilder();
        kernelBuilder.Services.ConfigureHttpClientDefaults(c => c.AddStandardResilienceHandler());
        var kernel = kernelBuilder
            .AddHuggingFaceTextEmbeddingGeneration("nomic-ai/nomic-embed-text-v1.5-GGUF/nomic-embed-text-v1.5.Q8_0.gguf",
            new Uri("http://localhost:1234/v1"),
            apiKey: "lm-studio",
            serviceId: null)
            .Build();

        var embeddingGenerator = kernel.GetRequiredService<ITextEmbeddingGenerationService>();
        var memoryBuilder = new MemoryBuilder();
        memoryBuilder.WithTextEmbeddingGeneration(embeddingGenerator);
        memoryBuilder.WithMemoryStore(new VolatileMemoryStore());
        var memory = memoryBuilder.Build();
        // Download a document and create embeddings for it
        string input = "What is an amphibian?";
        string[] examples = [ "What is an amphibian?",
                              "Cos'è un anfibio?",
                              "A frog is an amphibian.",
                              "Frogs, toads, and salamanders are all examples.",
                              "Amphibians are four-limbed and ectothermic vertebrates of the class Amphibia.",
                              "They are four-limbed and ectothermic vertebrates.",
                              "A frog is green.",
                              "A tree is green.",
                              "It's not easy bein' green.",
                              "A dog is a mammal.",
                              "A dog is a man's best friend.",
                              "You ain't never had a friend like me.",
                              "Rachel, Monica, Phoebe, Joey, Chandler, Ross"];
        for (int i = 0; i < examples.Length; i++)
            await memory.SaveInformationAsync("net7perf", examples[i], $"paragraph{i}");
        var embed = await embeddingGenerator.GenerateEmbeddingsAsync([input]);
        ReadOnlyMemory<float> inputEmbedding = (embed)[0];
        // Generate embeddings for each chunk.
        IList<ReadOnlyMemory<float>> embeddings = await embeddingGenerator.GenerateEmbeddingsAsync(examples);
        // Print the cosine similarity between the input and each example
        float[] similarity = embeddings.Select(e => TensorPrimitives.CosineSimilarity(e.Span, inputEmbedding.Span)).ToArray();
        similarity.AsSpan().Sort(examples.AsSpan(), (f1, f2) => f2.CompareTo(f1));
        Console.WriteLine("Similarity Example");
        for (int i = 0; i < similarity.Length; i++)
            Console.WriteLine($"{similarity[i]:F6}   {examples[i]}");
    }
}

At the line:

for (int i = 0; i < examples.Length; i++)
    await memory.SaveInformationAsync("net7perf", examples[i], $"paragraph{i}");

I get the following exception:

JsonException: The JSON value could not be converted to Microsoft.SemanticKernel.Connectors.HuggingFace.Core.TextEmbeddingResponse

Does anybody know what I'm doing wrong?

I've downloaded the following nuget packages into the project:

Id	Versions	ProjectName
Microsoft.SemanticKernel.Core	{1.15.0}	LocalLlmApp
Microsoft.SemanticKernel.Plugins.Memory	{1.15.0-alpha}	LocalLlmApp
Microsoft.Extensions.Http.Resilience	{8.6.0}	LocalLlmApp
Microsoft.Extensions.Logging	{8.0.0}	LocalLlmApp
Microsoft.SemanticKernel.Connectors.HuggingFace	{1.15.0-preview}	LocalLlmApp
Newtonsoft.Json	{13.0.3}	LocalLlmApp
Microsoft.Extensions.Logging.Console	{8.0.0}	LocalLlmApp

Answer 1

I found a solution to this problem thanks to Bruno Capuano's blog post about building a local RAG scenario using Phi-3 and SemanticKernel.

The code up to the string input = "What is an amphibian?"; line now looks like this:

    // Initialize the Semantic kernel
    IKernelBuilder kernelBuilder = Kernel.CreateBuilder();

    Kernel kernel = kernelBuilder
        .AddOpenAIChatCompletion(
                    modelId: "phi3",
                endpoint: new Uri("http://localhost:1234"),
                apiKey: "lm-studio")
        .AddLocalTextEmbeddingGeneration()
        .Build();

    // get the embeddings generator service
    var embeddingGenerator = kernel.Services.GetRequiredService<ITextEmbeddingGenerationService>();
    var memory = new SemanticTextMemory(new VolatileMemoryStore(), embeddingGenerator);

So although we're not using OpenAI we can still use the AddOpenAIChatCompletion method.

The AddLocalTextEmbeddingGeneration() method is from the SmartComponents.LocalEmbeddings.SemanticKernel Nuget package

I wrote a small console program with most of the examples from the blog posts. You can find it on github