Location>code7788 >text

Use SK for vector operations

Popularity：687 ℃/2025-03-01 16:44:47

I wish you a happy new year in 2025 first.
Judging from the LLM applications implemented in 2024, they are basically used in combination with RAG technology. Because most people and companies do not have the ability to turn fine-turn. Whether it is from the perspective of difficulty or cost, RAG technology is much more friendly.

In RAG (Retrieval-Augmented Generation), the meaning of vectors is to convert text data into high-dimensional vector representations for efficient similarity search and information retrieval. Specifically, the role of vectors in RAG includes:
Text embedding: convert text data (such as user query, document content) into vector representation. These vectors capture the semantic information of the text, making similar text closer in the vector space.
Similarity search: By calculating the distance between vectors (such as cosine similarity), you can quickly find the document vector that is most similar to the query vector, thereby achieving efficient information retrieval.
Enhanced generation: When generating text by generative models (such as GPT), the retrieved relevant document vectors are used as auxiliary information to improve the relevance and accuracy of the generated results.

Use SK to store and retrieve vectors

If you want to use RAG technology, it is basically inseparable from basic operations such as storing and retrieving vectors. Fortunately, SK has already packaged it all for us. Let's see how to use SK to play with vectors.

Define User Model class

Define the User Model class to describe the data structure. useVectorStoreRecordKeyAttributeIndicates the key field, useVectorStoreRecordDataAttributeIndicates the data field,VectorStoreRecordVectorIndicates the vector field.

        public class UserModel
    {
        [VectorStoreRecordKey]
        public string UserId { get; set; }

        [VectorStoreRecordData]
        public string UserName { get; set; }

        [VectorStoreRecordData]
        public string Hobby { get; set; }

        public string Description => $"{UserName}'s ID is {UserId} and hobby is {Hobby}";
        
        [VectorStoreRecordVector(1024, , )]
        public ReadOnlyMemory<float>? DescriptionEmbedding { get; set; }

    }

SK provides us withIVectorStoreinterface. In this way, various vector storage solutions are only necessary to implement this interface. SK provides us with many out-of-the-box libraries, such as: InMemory, Redis, Azure Cosmos, Qdrant, PG. Just install it through nuget.
Below we use Redis as a vector database to demonstrate it to you.

Use docker to install redis stack server

The default redis does not support vector search, we need to useredis/redis-stack-server:latestThis image.

docker run -d --name redis-stack-server -p 6379:6379 redis/redis-stack-server:latest

Initialize RedisVectorStore

 var vectorStore = new RedisVectorStore(
  ("localhost:6379").GetDatabase(),
  new() { StorageType =  });

Initialize collection

Create a collection to store user information. Collection can be considered as a table in a relational database.

  // init collection
   var collection = <string, UserModel>("ks_user");
   await ();

Initialize EmbeddingGenerationService

The following is to use the local ollama service to provide the embedded generation service. This service is at the heart of all text to vectors.

 // init embedding serivce
    var ollamaApiClient = new OllamaApiClient(new Uri(ollamaEndpoint), modelName);
    var embeddingGenerator = ();

Vector CRUD

The following code demonstrates how to convert the Description field of User into a vector and perform the most basic Insert, Update, Delete, and Get operations.

// init user infos and vector
var users = ();
 foreach (var user in users)
 {
      = await ();
 }

// insert or update
foreach (var user in users)
{
    await (user);           
}

// get
var alice = await ("1");
();
var all = ((x=>));
await foreach(var user in all)
{
    ();
}

// delete
await ("1");

Vector Search

The following demonstrates how to perform a vector awareness search. First, generate the text of the problem in a vector, and then use this vector for searching. When searching, you can configure matching fields and take the first few results.

// search
var vectorSearchOptions = new VectorSearchOptions
{
    VectorPropertyName = nameof(),
    Top = 3
};
var query = await ("Who hobby is swimming?");
var searchResult = await (query,vectorSearchOptions);
await foreach (var user in )
{
    ();
    ();
}

Summarize

Above we demonstrate how to vectorize the data model and cooperate with redis to perform basic operations on CRUD. At the same time, vectorized search of text problems, that is, similar searches are also demonstrated. Although the above demonstration runs with redis, SK also provides us with a lot of choices, and you can quickly choose your favorite vector database for storage. For example: Azure Cosmos, Qdrant, PG, SQLite, etc. Okay, there is nothing to say. I hope this article can help everyone learn SemanticKernel, thank you.

Sample code has been uploaded to github
/kklldog/SKLearning

<Previous Next>