[Bug/Question] GetNearestMatchesAsync missing on IVectorStoreVectorCollection (Semantic Kernel v1.51.0 + Sqlite 1.51.0-preview + Google 1.52.0-alpha)

[roytarak]

Summary

I'm working on a C# service that uses Semantic Kernel’s SqliteVectorStore together with GoogleAIEmbeddingGenerator (models/embedding-001) to store and search embeddings.

I’m running into a compilation issue when attempting to call GetNearestMatchesAsync on a SqliteVectorStoreRecordCollection.

Environment Details

NuGet Package Versions (from dotnet list package):

Package	Version
Microsoft.SemanticKernel	1.51.0
Microsoft.SemanticKernel.Connectors.Sqlite	1.51.0-preview
Microsoft.SemanticKernel.Connectors.Google	1.52.0-alpha
Microsoft.Extensions.VectorData.Abstractions	9.0.0-preview.1.25229.1

Environment:

Windows Server 2022
.NET SDK 8.0.401
Visual Studio 2022

Minimal Reproducible Code:

using Microsoft.SemanticKernel.Connectors.Google;
using Microsoft.SemanticKernel.Connectors.Sqlite;
using Microsoft.Extensions.VectorData;
using Microsoft.Extensions.VectorData.Core;
using System;
using System.Collections.Generic;
using System.Linq;
using System.Threading.Tasks;

namespace VectorSearchRepro
{
public class SqliteVectorRecord : IVectorStoreRecord
{
public string Id { get; set; } = string.Empty;
public string Text { get; set; } = string.Empty;
public ReadOnlyMemory Vector { get; set; }

    public ReadOnlyMemory<float> GetVector() => Vector;
    public void SetVector(ReadOnlyMemory<float> vector) => Vector = vector;
}

public class Program
{
    private static SqliteVectorStore _vectorStore;
    private static IVectorStoreVectorCollection<string, SqliteVectorRecord> _collection;
    private static GoogleAIEmbeddingGenerator _generator;

    public static async Task Main()
    {
        string dbPath = "semantic_vectors.db";
        string geminiApiKey = Environment.GetEnvironmentVariable("GEMINI_API_KEY");

#pragma warning disable SKEXP001
#pragma warning disable SKEXP0070
_generator = new GoogleAIEmbeddingGenerator(
modelId: "models/embedding-001",
apiKey: geminiApiKey);
#pragma warning restore SKEXP0070
#pragma warning restore SKEXP001

        _vectorStore = new SqliteVectorStore(dbPath);
        _collection = new SqliteVectorStoreRecordCollection<string, SqliteVectorRecord>(
            _vectorStore, "agreements");

        var chunks = new List<string> { "Confidentiality clause text." };

        var result = await _generator.GenerateAsync(chunks);
        var embedding = result.Data.First().Vector;

        var record = new SqliteVectorRecord
        {
            Id = "1",
            Text = chunks[0],
            Vector = embedding
        };

        await _collection.UpsertAsync(record);

        // ❌ Error here:
        var results = await _collection.GetNearestMatchesAsync(
            embedding.ToArray(), limit: 3);

        Console.WriteLine(results);
    }
}

}

Error Message: 'IVectorStoreVectorCollection<string, SqliteVectorRecord>' does not contain a definition for 'GetNearestMatchesAsync'
and no accessible extension method 'GetNearestMatchesAsync' accepting a first argument of type
'IVectorStoreVectorCollection<string, SqliteVectorRecord>' could be found
(are you missing a using directive or an assembly reference?)

What I’ve Tried

Added using Microsoft.Extensions.VectorData;
Added using Microsoft.Extensions.VectorData.Core;
Verified _collection type is SqliteVectorStoreRecordCollection<string, SqliteVectorRecord>
Checked Semantic Kernel and VectorData assemblies for extension methods
No overload or extension method for nearest neighbor search appears to exist

Question

In Semantic Kernel v1.51.0 (with VectorData 9.x):
Where has the GetNearestMatchesAsync method been moved?
Which interface or class now exposes nearest-neighbor vector search for SqliteVectorStore collections?
Is there a new API (e.g., IVectorSearchCollection, IVectorQueryExecutor) that replaces this method?
If the API was refactored between 1.51.0-preview and 1.52.0-alpha, could you please point to the relevant example or PR?

Goal

I simply want to:
Generate embeddings using GoogleAIEmbeddingGenerator (models/embedding-001)
Store them in a local SQLite vector store
Perform semantic similarity search (kNN)
using Semantic Kernel’s built-in vector connectors.

Expected Behavior
_collection.GetNearestMatchesAsync() should return the top-N semantically similar text records.

Actual Behavior
Compilation fails — the method is no longer defined on IVectorStoreVectorCollection.

[Samielakkad]

GetNearestMatchesAsync belongs to the older IMemoryStore API. It is not the search method on the Vector Data collection used by SqliteVectorStoreRecordCollection in dotnet-1.51.0.

At that exact tag, IVectorStoreRecordCollection<TKey, TRecord> inherits IVectorSearch<TRecord>, whose two search methods are:

• SearchEmbeddingAsync(...) when you already generated the embedding;
• SearchAsync(...) when the collection has an embedding generator configured and should generate it internally.

The 1.51.0 source is here:

• interface: https://github.com/microsoft/semantic-kernel/blob/dotnet-1.51.0/dotnet/src/Connectors/VectorData.Abstractions/VectorSearch/IVectorSearch.cs
• SQLite implementation: https://github.com/microsoft/semantic-kernel/blob/dotnet-1.51.0/dotnet/src/Connectors/Connectors.Memory.Sqlite/SqliteVectorStoreRecordCollection.cs
• working sample: https://github.com/microsoft/semantic-kernel/blob/dotnet-1.51.0/dotnet/samples/Concepts/Memory/VectorStore_VectorSearch_Simple.cs

Because your code already creates the embedding, the direct replacement is:

private static IVectorStoreRecordCollection<string, SqliteVectorRecord> _collection;

// ... create/upsert as before

var results = await _collection
    .SearchEmbeddingAsync(embedding, top: 3)
    .ToListAsync();

foreach (var match in results)
{
    Console.WriteLine($"{match.Score}: {match.Record.Text}");
}

The important type change is IVectorStoreRecordCollection<...>. Your declared IVectorStoreVectorCollection<...> comes from a different preview generation of Microsoft.Extensions.VectorData, so it does not expose the 1.51 search contract.

I would also align the package versions. The example mixes Semantic Kernel 1.51.0, SQLite 1.51.0-preview, Google 1.52.0-alpha, and a separately pinned Microsoft.Extensions.VectorData.Abstractions preview. With prerelease packages, that can compile against incompatible interface generations. Remove the direct abstractions reference unless you specifically need it, then use the version brought in by the chosen connector set, or move all related packages to one compatible release line.

VectorizedSearchAsync also exists at 1.51, but it is already marked obsolete there in favor of SearchEmbeddingAsync, so it is not the best migration target.