Local LLM Server for Semantic Kernel .NET Developer

如果你使用中文，请访问这里

Support Semantic-Kernel 1.0.1

Or you are using Semantic Kernel's Hugging Face http server as your local LLM service , but based on the inaccessibility of hugging face in mainland China and management reasons, I tried to reconstruct the project. At this stage, it is adapted for macOS and Linux environments.

At this stage, the implementation of ChatCompletion and Embedding has been completed.

ChatCompletion is adapted to LLM baichu2, ChatGLM3, Microsoft-phi 1.5

Embeddings adapted to LLM jina-embeddings (English Embedding model), text2vec-large-chinese (Chiense Embedding model), bge-*-v1.5(Chinese/English Embedding model)

Samples

0. download your LLM firstly and using pip to install python library

pip install -r requirement.txt

1. .env config your ChatCompletion and Embedding model location

CHAT_COMPLETION_URL = 'Your chat completion model location'
EMBEDDING_URL = 'Your embeddings model location'

2. Start your Local LLM Http Server

python local_llm_service.py

3. Add Microsoft.SemanticKernel, Microsoft.SemanticKernel.Connectors.AI.HuggingFace, Microsoft.SemanticKernel.Connectors.Memory.Qdrant(You can choose different vector database) packages

4. Initialization endpoint for chatcompletion, embeddings, and qdrant

string chat_endpoint = "http://localhost:5002/v1/chat/completions";
string embeddings_endpoint = "http://localhost:5002/v1/embeddings";
string qdrant_endpoint = "http://localhost:6333";

5. Sample 1 - ChatCompletion

using Microsoft.SemanticKernel;
using Microsoft.SemanticKernel.Connectors.Memory.Qdrant;
using Microsoft.SemanticKernel.Plugins.Memory;
using Microsoft.SemanticKernel.Connectors.AI.HuggingFace.TextEmbedding;

#pragma warning disable SKEXP0020

Kernel kernel = new KernelBuilder()
            .AddHuggingFaceTextGeneration(
                model: "baichuan2",
                endpoint: chat_endpoint)
            .Build();

var questionAnswerFunction = kernel.CreateFunctionFromPrompt("问: {{$input}} 答:");

var result = await kernel.InvokeAsync(questionAnswerFunction, new("介绍一下自己"));

result.GetValue<string>()

6. Sample 2 - Embeddings

#pragma warning disable SKEXP0052
#pragma warning disable CS1061
#pragma warning disable SKEXP0011
#pragma warning disable SKEXP0026

#pragma warning disable SKEXP0020

var qdrantMemoryBuilder = new MemoryBuilder();

var hfembeddings = new HuggingFaceTextEmbeddingGeneration("text2veccn", embeddings_endpoint);

qdrantMemoryBuilder.WithTextEmbeddingGeneration(hfembeddings);
qdrantMemoryBuilder.WithQdrantMemoryStore(qdrant_endpoint, 1024);

var builder = qdrantMemoryBuilder.Build();

string MemoryCollectionName = "text2vecdemo";


await builder.SaveInformationAsync(MemoryCollectionName, id: "id1", text: "我是卢建晖");
await builder.SaveInformationAsync(MemoryCollectionName, id: "id2", text: "卢建晖是微软云技术布道师");
await builder.SaveInformationAsync(MemoryCollectionName, id: "id3", text: "卢建晖从 2012 年到 2020 年是微软最有价值专家");
await builder.SaveInformationAsync(MemoryCollectionName, id: "id4", text: "卢建晖是人工智能讲师");

var searchResults =  builder.SearchAsync(MemoryCollectionName, "你认识卢建晖吗", limit: 3, minRelevanceScore: 0.6);

await foreach (var item in searchResults)
{
    Console.WriteLine(item.Metadata.Text + " : " + item.Relevance);
}

If you want to get English text embedding and chat completion , please click here

🍔🍔🍔🍔🍔🍔🍔🍔🍔 More functions will be added in the future