mirror of
https://github.com/microsoft/semantic-kernel.git
synced 2026-10-12 02:19:06 +00:00
Abby Harrison
1e72692591
### Motivation and Context Existing memory interfaces are overly complicated, leading to confusion about how to implement a custom Memory Store. This PR condenses `IMemoryStore` + `IEmbeddingIndex` + `IDataStore `-> **`IMemoryStore`** and eliminates the use of generics spread by the Embedding struct. It makes the codebase much easier to read and IMemoryStore easier to implement. As part of the effort, SqliteMemoryStore was properly implemented (it was not previously). Additionally, this PR moves the MemoryStore 'skills' to the connectors namespace. Addresses several points called out in https://github.com/microsoft/semantic-kernel/issues/202 ### Description - `QdrantMemoryStore` was moved from Skills to Connectors. Unit tests were added for the memory store object. Created an IQdrantVectorDbClient interface to better enable testing of the memory store through mocks. - Modified QdrantVectorDbClient calls to default batched calls to reduced the amount of HTTP traffic when possible. `GetVectorById -> GetVectorsById`, `UpsertVector -> UpsertVectors`, `DeleteVector -> DeleteVectors`. - Generally, for a vector store, the unique id passed by the user to associate with a saved memory will become the indexed ID for the vector store as well. However, Qdrant enforces a user-provided GUID or hex string as the index key. Added a nullable key field to `MemoryRecord` to support vector storage services that act similarly. Depending on the backing storage, this key is not necessary to pass to subsequent calls. `SemanticTextMemory` does not utilized this parameter. For now, this is only for direct interaction with the `QdrantMemoryStore`. **Note**: It is understood that Qdrant will be enabling service-side GUID generation in the near future, significantly reducing the stress on the client to produce unique values for every data entry. - Eliminated the dependency on the core kernel in the connectors by replacing `Embeddings` struct references with `IEnumerable<float>`. - `SqliteMemoryStore` was moved from Skills to Connectors. `SqliteMemoryStore` now properly implements `IMemoryStore` and the unit tests were updated to reflect this fact. - The generic `TEmbedding` was removed from all Kernel and Connector files. **If a memory store expects a non-FP32 vector, it is currently up to the `IMemoryStore` implementation to perform the necessary conversions.** - `DataEntry` is now a base class that enforces nullable Key and Timestamp fields. It is not strictly used for anything other than MemoryRecord. It may serve useful for connecting to data stores that are for structures other than embeddings, or vector stores that calculate embeddings server-side. - Removed existing collection checks for `Upsert` calls. Depending on the implementation, this has the potential to drastically increase HTTP traffic or local compute resources when making batch calls. The call has been moved to SemanticTextMemory for now to avoid additional breaking changes, but will be addressed in a better fashion in a follow-up PR. - Added a very simple example to `KernelSyntaxExamples `on how to create a read-only `IMemoryStore` that imports vectors from a json string. - Renamed `VectorDbException` -> `QdrantMemoryException`. Ensured that these exceptions can only be thrown by the `QdrantMemoryStore` and can accept an inner exception as a paramter. Additional small edits: - removed unnecessary namespace from KernelSyntaxExamples Example21 and Example22 - Removed unused references in `RestApiOperations.cs` and `JsonPathSkill.cs` ---------------- OUT OF SCOPE, COMING SOON ---------------- - The addition of a value string to MemoryRecord. This is very important for metadata flexibility, but it will come in a follow-up PR. - The ability to optionally return the vector to the user. In many cases, the vector is not important for get/search calls, thus it makes no sense to pass such large amounts of data around the kernel application. - Disambiguation of SemanticTextMemory interfaces, classes, and constructors. - Unit tests for QdrantVectorDbClient. Qdrant will be making breaking changes to their API very soon. TBD whether integration tests will be added as well.
40 lines
1.2 KiB
C#
40 lines
1.2 KiB
C#
// Copyright (c) Microsoft. All rights reserved.
|
|
|
|
using System.IO;
|
|
using System.Text.Json;
|
|
using Microsoft.Extensions.Configuration;
|
|
using Microsoft.Extensions.DependencyInjection;
|
|
using Microsoft.Extensions.Hosting;
|
|
using Microsoft.SemanticKernel.Memory;
|
|
|
|
namespace KernelHttpServer;
|
|
|
|
public static class Program
|
|
{
|
|
public static void Main()
|
|
{
|
|
var host = new HostBuilder()
|
|
.ConfigureFunctionsWorkerDefaults()
|
|
.ConfigureAppConfiguration(configuration =>
|
|
{
|
|
var config = configuration.SetBasePath(Directory.GetCurrentDirectory())
|
|
.AddJsonFile("local.settings.json", optional: true, reloadOnChange: true);
|
|
|
|
var builtConfig = config.Build();
|
|
})
|
|
.ConfigureServices(services =>
|
|
{
|
|
services.AddSingleton<IMemoryStore>(new VolatileMemoryStore());
|
|
|
|
// return JSON with expected lowercase naming
|
|
services.Configure<JsonSerializerOptions>(options =>
|
|
{
|
|
options.PropertyNamingPolicy = JsonNamingPolicy.CamelCase;
|
|
});
|
|
})
|
|
.Build();
|
|
|
|
host.Run();
|
|
}
|
|
}
|