Vector search

Azure Cosmos DB now offers preview support for vector similarity search. Vector search is a fundamental part of some application types, include AI, semantic search and others. The Cosmos DB support for vector search allows storing your data and vectors, and performing your queries in a single database, which can considerably simplify your architecture and remove the need for an additional, dedicated vector database solution in your stack. To learn more about Cosmos DB vector search, see the documentation.

To use vector search, you must first enroll in the preview feature. Then, define the vector policies on your container, which determine which JSON property in your documents contain vectors, various vector-related information for those properties (dimensions, data type, distance function).

Once your container is properly set up, add a vector property to your model in the path you defined in the container policy, and configure it with EF as a vector:

public class Blog
{
    ...

    public float[] Vector { get; set; }
}

public class BloggingContext
{
    ...

    protected override void OnModelCreating(ModelBuilder modelBuilder)
    {
        modelBuilder.Entity<Blog>()
            .Property(b => b.Embeddings)
            .IsVector(DistanceFunction.Cosine, dimensions: 1536);
    }
}

At this point your model is configured. Insertion of vector data is done just like any other data type with EF:

float[] vector = /* generate vector data from text, image, etc. */
context.Add(new Blog { Vector = vector });
await context.SaveChangesAsync();

Finally, use the EF.Functions.VectorDistance() function in LINQ queries to perform vector similarity search:

float[] anotherVector = /* generate vector data from text, image, etc. */
var blogs = await context.Blogs
    .OrderBy(s => EF.Functions.VectorDistance(s.Vector, anotherVector))
    .Take(5)
    .ToListAsync();

This will returns the top five Blogs, based on the similarity of their Vector property and the externally-provided anotherVector data.