Reimplement ToolDescriptionEvaluator VectorDB on Microsoft.Extensions.VectorData

[Copilot]

The ToolDescriptionEvaluator's VectorDB was a bespoke in-memory store with a custom API. Building it on the Microsoft.Extensions.VectorData abstractions lets it be swapped for any provider (Azure AI Search, Cosmos DB, Redis, …) with little to no consumer changes, and unblocks Microsoft.Extensions.DataIngestion for chunking/processing.

Changes

• Abstractions: Derived VectorDB from VectorStore and added InMemoryVectorStoreCollection : VectorStoreCollection<string, Entry>. The original brute-force cosine-similarity search, sorted-key storage, and parallel divide-and-conquer query path are preserved behind the standard UpsertAsync/SearchAsync/GetAsync/DeleteAsync surface.

• Record schema: Entry is now annotated for vector data:

  public sealed record Entry(
      [property: VectorStoreKey] string Id,
      [property: VectorStoreData] object? Metadata,
      [property: VectorStoreVector(3072, DistanceFunction = DistanceFunction.CosineSimilarity)]
      ReadOnlyMemory<float> Vector);

• Pluggable metric retained: IDistanceMetric/CosineSimilarity/DotProduct kept, now span-based for zero-alloc TensorPrimitives calls.

• Consumer (Program.cs): Resolves the collection from the store and uses the async API; search results are read as VectorSearchResult<Entry> (.Record/.Score).

• Packaging: Added Microsoft.Extensions.VectorData.Abstractions 10.0.1 via central package management (aligned with the repo's pinned Microsoft.Extensions.AI.Abstractions 10.3.0).

Notes

• GetAsync(Expression<...>) intentionally throws NotSupportedException — implementing it requires Expression.Compile(), which is AOT-incompatible; the tool never calls it.
• Parallel search now caps the merged result at top (previously up to 2*TopK); consumers only ever read the first N ≤ top, so ranking is unchanged.

[azure-pipelines]

Azure Pipelines:
Successfully started running 1 pipeline(s).
There may be pipelines that require an authorized user to comment /azp run to run.