EF Core 11 Adds Native SQL Server Vector Search with DiskANN Indexes

EF Core 10 introduced EF.Functions.VectorDistance() for computing exact distances between embeddings in LINQ queries. That works, but exact search over millions of rows is expensive. EF Core 11 Preview 2 closes the gap by supporting SQL Server 2025’s approximate vector search: DiskANN indexes and the VECTOR_SEARCH() table-valued function, all wired up through your DbContext.

Setting up a vector index

Declare the index in OnModelCreating with the distance metric you want (cosine, dot product, or Euclidean):

protected override void OnModelCreating(ModelBuilder modelBuilder)
{
    modelBuilder.Entity<Blog>()
        .HasVectorIndex(b => b.Embedding, "cosine");
}

When you add a migration, EF generates the CREATE VECTOR INDEX DDL targeting SQL Server 2025’s DiskANN engine. The index lives alongside your regular B-tree and full-text indexes, managed through the same migration pipeline.

Querying with VectorSearch()

Once the index exists, use the new VectorSearch() extension method on your DbSet:

float[] queryEmbedding = GetEmbeddingForQuery("distributed caching");

var results = await context.Blogs
    .VectorSearch(b => b.Embedding, "cosine", queryEmbedding, topN: 5)
    .ToListAsync();

This translates to SQL Server’s VECTOR_SEARCH() table-valued function, which performs an approximate nearest-neighbor lookup against the DiskANN index. The topN parameter caps how many results come back.

The return type is VectorSearchResult<TEntity>, which exposes both the entity and the computed distance:

var results = await context.Blogs
    .VectorSearch(b => b.Embedding, "cosine", queryEmbedding, topN: 10)
    .Select(r => new { r.Value.Name, r.Distance })
    .ToListAsync();

Exact vs. approximate: when to use which

VectorDistance() from EF Core 10 still works and gives you exact results. Use it when the dataset is small or precision matters more than latency. VectorSearch() with a DiskANN index trades a small amount of recall accuracy for dramatically better throughput on large tables.

In practice, most RAG and recommendation workloads want the approximate path. If you previously offloaded vector search to a dedicated database (Qdrant, Pinecone, pgvector), this brings it back into the SQL Server you are already running, with EF Core managing the schema.

Requirements

This feature targets SQL Server 2025, which introduced DiskANN vector indexes. The VECTOR_SEARCH() function and the related CREATE VECTOR INDEX syntax are experimental in SQL Server at the time of writing, so expect changes. The EF Core APIs mirror that experimental status.

For full setup details, see the EF Core vector search documentation.