Redis
redis-vl-dotnet Documentation redis-vl-dotnet

redis-vl-dotnet

Microsoft.Extensions.VectorData Connector

RedisVL.Connectors.VectorData exposes a RedisVL SearchIndex as a Microsoft.Extensions.VectorData (MEVD) VectorStore / VectorStoreCollection<TKey, TRecord>. Because Semantic Kernel’s memory and vector-store connectors are built on MEVD, this single package is consumable from Semantic Kernel and the broader Microsoft.Extensions.AI ecosystem.

This is the .NET analog of the Java library’s LangChain4J adapters (RedisVLEmbeddingStore, RedisVLContentRetriever, RedisVLChatMemoryStore, and the filter mapper).

Package contents

  • RedisVLVectorStore — a MEVD VectorStore over a StackExchange.Redis IDatabase

  • RedisVLCollection<TKey, TRecord> — a MEVD VectorStoreCollection backed by a JSON-storage SearchIndex

  • RedisVLCollectionOptions — index name, key prefix, and an explicit VectorStoreCollectionDefinition

  • RedisVLChatMessageStore (namespace RedisVL.Connectors.VectorData.ChatMemory) — a chat-history store that bridges Microsoft.Extensions.AI ChatMessage to the MessageHistory workflow

Record model

The connector maps a record type onto a JSON-backed search index using the standard MEVD attributes, or an explicit VectorStoreCollectionDefinition passed through RedisVLCollectionOptions.Definition:

  • [VectorStoreKey] — the document key. Only string keys are supported.

  • [VectorStoreData(IsIndexed = true)] — indexed as a TAG field (strings, enums, booleans, string collections) or a sortable NUMERIC field (numeric types).

  • [VectorStoreData(IsFullTextIndexed = true)] — indexed as a TEXT field (strings only).

  • [VectorStoreData] without indexing flags — stored in JSON but not added to the index.

  • [VectorStoreVector(dims, DistanceFunction = …​, IndexKind = …​)] — a VECTOR field. ReadOnlyMemory<float>/float[]/Embedding<float> map to FLOAT32; ReadOnlyMemory<double>/double[] map to FLOAT64.

Field (and JSON) names follow System.Text.Json web defaults — camel-cased property names, or a [JsonPropertyName] override.

Distance functions map to RedisVL metrics: cosine → Cosine, dot-product → InnerProduct, Euclidean → L2. Index kinds map Flat → Flat and Hnsw/Dynamic → Hnsw.

Usage

var store = new RedisVLVectorStore(database);
var movies = store.GetCollection<string, Movie>("movies");

await movies.EnsureCollectionExistsAsync();
await movies.UpsertAsync(catalog);

// Vector search with a LINQ metadata pre-filter (translated to a RedisVL FilterExpression).
await foreach (var result in movies.SearchAsync(
    queryVector,
    top: 5,
    new VectorSearchOptions<Movie> { Filter = m => m.Genre == "scifi" && m.Year >= 1990 }))
{
    Console.WriteLine($"{result.Record.Title} ({result.Score})");
}

Filter translation

LINQ predicates on VectorSearchOptions.Filter and GetAsync(filter, top) are translated to RedisVL FilterExpression values. Supported forms include ==, !=, <, ⇐, >, >=, &&, ||, !, collection Contains (tag membership and IN), and captured-variable evaluation.

Limitations

  • Search input must be a raw vector (ReadOnlyMemory<float>, float[], Embedding<float>, or the double equivalents). Passing a string for server-side or generator-based embedding is not yet supported.

  • Only string keys are supported.

  • Dynamic (dictionary-based) collections are not supported; use GetCollection<TKey, TRecord>.

  • Records are always materialized in full, so IncludeVectors = false is treated as best-effort.

Semantic Kernel interop

Because Semantic Kernel’s vector-store and text-search abstractions are built on MEVD, the connector plugs directly into SK’s VectorStoreTextSearch<TRecord> for retrieval-augmented generation:

var collection = new RedisVLVectorStore(database).GetCollection<string, Movie>("movies");
var textSearch = new VectorStoreTextSearch<Movie>(collection, embeddingGenerator, stringMapper, resultMapper);
var results = await textSearch.GetTextSearchResultsAsync("a natural-language question");
var plugin = textSearch.CreateWithGetTextSearchResults("RedisVLSearch"); // expose to a Kernel

The package targets Microsoft.Extensions.VectorData.Abstractions 10.1.0 as its floor to stay binary-compatible with the Semantic Kernel 1.x line (a newer MEVD added a class constraint to VectorSearchResult<T> that SK 1.77 cannot load). NuGet unifies upward when a consumer brings a newer compatible MEVD.

Example workflows

  • /examples/VectorDataConnectorExample — the full MEVD flow: store creation, OpenAI-embedded upsert, fetch-by-key, filtered vector search, filtered retrieval, and the chat-memory store.

  • /examples/SemanticKernelConnectorExample — Semantic Kernel consuming the connector through VectorStoreTextSearch<TRecord>, with OpenAI embeddings.

Both examples require OPENAI_API_KEY and embed with OpenAI (text-embedding-3-small by default) through a Microsoft.Extensions.AI IEmbeddingGenerator.

Run them from the repository root:

dotnet run --project examples/VectorDataConnectorExample/VectorDataConnectorExample.csproj
dotnet run --project examples/SemanticKernelConnectorExample/SemanticKernelConnectorExample.csproj

Validation references

  • tests/RedisVL.Tests/Connectors/VectorData/RedisVLRecordModelTests.cs covers schema mapping

  • tests/RedisVL.Tests/Connectors/VectorData/RedisVLFilterTranslatorTests.cs covers LINQ filter translation

  • tests/RedisVL.Tests/Connectors/VectorData/RedisVLVectorStoreIntegrationTests.cs covers the collection round-trip

  • tests/RedisVL.Tests/Connectors/VectorData/RedisVLChatMessageStoreIntegrationTests.cs covers the chat-memory store