RedisVL.Vectorizers.Onnx 0.1.0

redis-vl-dotnet

redis-vl-dotnet is a .NET-native Redis Vector Library for Redis Search and vector workloads. It gives you a fluent, strongly-typed API over Redis Query Engine so you can define schemas, index HASH or JSON documents, and run vector, full-text, filter, hybrid, and aggregation queries — plus higher-level building blocks for GenAI apps such as semantic caching, semantic routing, embeddings caching, and semantic message history.

It is the .NET sibling of redis-vl-python, built on top of StackExchange.Redis.

Preview: redis-vl-dotnet is pre-1.0 and under active development. APIs may change between releases; pin an exact package version and review the CHANGELOG before upgrading.

Features

• Schema and index management — define indexes in code or load them from YAML, create/drop/list indexes, and reconnect to existing ones.
• Document storage — store and fetch strongly-typed records over HASH or JSON, with partial updates.
• Query types — nearest-neighbor VectorQuery, MultiVectorQuery, vector range queries, FilterQuery, TextQuery, AggregationQuery, and native FT.HYBRID HybridSearchQuery (Redis 8.4+).
• Fluent filters — compose tag, text, numeric, and geo predicates with &, |, and negation.
• Vector indexing options — FLAT, HNSW, and SVS-VAMANA algorithms, with LVQ / LeanVec compression.
• GenAI extensions — SemanticCache, SemanticRouter, EmbeddingsCache, and semantic message history.
• Pluggable vectorizers and rerankers — OpenAI, Cohere, Hugging Face, local ONNX, and Microsoft.Extensions.AI adapters, shipped as separate packages.
• Ecosystem connectors — a Microsoft.Extensions.VectorData / Semantic Kernel vector-store connector.

Requirements

• A supported .NET runtime: the packages multi-target .NET 8 (LTS), .NET 9, and .NET 10 (LTS).
• A Redis deployment with the Redis Query Engine (RediSearch) module — Redis 8 or Redis Cloud.
• RedisJSON for JSON-backed document workflows (included in Redis 8).
• Redis 8.4+ for native FT.HYBRID hybrid search.

You can start a local Redis with everything enabled via Docker:

docker run -d --name redis -p 6379:6379 redis:8

Installation

Install the core package from NuGet:

dotnet add package RedisVL

Add extension packages only as you need them:

Quick start

The core flow is: connect to Redis, define a schema, create the index, load documents, then query.

using RedisVL.Filters;
using RedisVL.Indexes;
using RedisVL.Queries;
using RedisVL.Schema;
using StackExchange.Redis;

// 1. Connect to Redis (StackExchange.Redis).
using var redis = await ConnectionMultiplexer.ConnectAsync("localhost:6379");
var database = redis.GetDatabase();

// 2. Define a schema: an index over the "movies:" keyspace with a 2-dim vector field.
var schema = new SearchSchema(
    new IndexDefinition("movies-idx", "movies:", StorageType.Hash),
    [
        new TagFieldDefinition("genre"),
        new TextFieldDefinition("title"),
        new VectorFieldDefinition(
            "embedding",
            new VectorFieldAttributes(
                VectorAlgorithm.Hnsw,
                VectorDataType.Float32,
                VectorDistanceMetric.Cosine,
                dimensions: 2,
                m: 16,
                efConstruction: 200))
    ]);

// 3. Create the index (dropping any existing one with the same name).
var index = new SearchIndex(database, schema);
await index.CreateAsync(new CreateIndexOptions(overwrite: true, dropExistingDocuments: true));

// 4. Load documents. Here we write HASHes directly; in real apps embeddings come from a vectorizer.
static byte[] EncodeFloat32(float[] vector)
{
    var bytes = new byte[vector.Length * sizeof(float)];
    Buffer.BlockCopy(vector, 0, bytes, 0, bytes.Length);
    return bytes;
}

await database.HashSetAsync("movies:heat", new HashEntry[]
{
    new("title", "Heat"),
    new("genre", "crime"),
    new("embedding", EncodeFloat32(new[] { 1f, 0f }))
});

// 5. Run a nearest-neighbor query, filtered to the "crime" genre.
var query = VectorQuery.FromFloat32(
    fieldName: "embedding",
    vector: new[] { 1f, 0f },
    topK: 3,
    filter: Filter.Tag("genre").Eq("crime"),
    returnFields: ["title"],
    scoreAlias: "distance");

var results = await index.SearchAsync(query);
foreach (var doc in results.Documents)
{
    Console.WriteLine($"{doc.Values["title"]} (distance={doc.Values["distance"]})");
}

Redis expects vector fields as raw little-endian float32 bytes, hence the EncodeFloat32 helper. See VectorSearchExample for a complete, runnable version.

JSON documents

Prefer JSON storage when you want to work with strongly-typed POCOs and JSONPath partial updates:

var schema = SearchSchema.FromYamlFile("schema.yaml");     // or build in code
var index = new SearchIndex(database, schema);
await index.CreateAsync(new CreateIndexOptions(overwrite: true));

var movies = new[]
{
    new Movie("movie-1", "Heat", 1995, "crime"),
    new Movie("movie-2", "Arrival", 2016, "science-fiction")
};
await index.LoadJsonAsync(movies);

var arrival = await index.FetchJsonByIdAsync<Movie>("movie-2");

// Filter + text queries, projected onto a type
var query = new FilterQuery(
    Filter.And(
        Filter.Tag("genre").Eq("science-fiction"),
        Filter.Numeric("year").GreaterThan(1980)),
    returnFields: ["title", "year", "genre"]);
var scienceFiction = await index.SearchAsync<Movie>(query);

Generating embeddings with a vectorizer

Vectorizer packages implement IBatchTextVectorizer, so you can turn text into embeddings and feed them to any query or cache:

using RedisVL.Vectorizers.OpenAI;

var vectorizer = new OpenAiTextVectorizer(
    model: "text-embedding-3-small",
    apiKey: Environment.GetEnvironmentVariable("OPENAI_API_KEY")!,
    new OpenAiVectorizerOptions { Dimensions = 256 });

float[] embedding = await vectorizer.VectorizeAsync("A linguist decodes an alien language.");
IReadOnlyList<float[]> batch = await vectorizer.VectorizeAsync(["first prompt", "second prompt"]);

Semantic cache

SemanticCache stores prompt/response pairs keyed by embedding similarity, so repeated (or paraphrased) prompts return a cached answer instead of hitting your LLM:

using RedisVL.Caches;
using RedisVL.Schema;

var cache = new SemanticCache(
    database,
    new SemanticCacheOptions(
        "semantic-cache-example",
        new VectorFieldAttributes(
            VectorAlgorithm.Flat, VectorDataType.Float32, VectorDistanceMetric.Cosine, dimensions: 3),
        distanceThreshold: 0.2d,
        timeToLive: TimeSpan.FromMinutes(10)));
await cache.CreateAsync();

await cache.StoreAsync(
    prompt: "How do I reset my password?",
    response: "Open Settings > Security > Reset password and follow the email link.",
    embedding: new[] { 1f, 0f, 0f });

// A near-identical embedding returns the cached response.
var hit = await cache.CheckAsync(prompt: "How do I reset my password?", embedding: new[] { 1f, 0f, 0f });
Console.WriteLine(hit?.Response);

Connecting to a cluster

RedisVL operates on a StackExchange.Redis IDatabase, so any connection you can build works. For cluster topologies, the RedisConnectionFactory helpers normalize seed-node lists and build ConfigurationOptions for you:

• ConnectClusterAsync(...) — cluster discovery
• CreateClusterOptions(...) — parsed option builder

Examples

The examples/ directory contains runnable projects for every feature area. Run any of them from the repository root, for example:

dotnet run --project examples/VectorSearchExample/VectorSearchExample.csproj

Highlights:

Documentation

Full documentation lives in the Antora docs site source:

• Overview
• Getting Started — installation, connection topologies, environment variables
• Core Features — schema, indexes, documents, queries
• Extensions — vectorizers, rerankers, connectors
• Testing

Build the docs locally from the repository root:

npm install
npm run docs:validate

Contributing

Issues and pull requests are welcome. The solution file is redis-vl-dotnet.sln; build and test with the standard dotnet CLI:

dotnet build redis-vl-dotnet.sln
dotnet test

Integration tests need a reachable Redis with the Query Engine enabled. See Testing for details.