Vector Search Support (Work in Progress)

Introduction

Starting with package version 2.1 (beta) it is possible to utilize the vector support present in RediSearch.

Defining an Index

RediSearchClient supports the creation of indexes by extending the existing index creation builder classes by introducing the Vector method. This new extension method is available on both JSON and HASH based indexes and supports HNSW and FLAT vector indexes.

The following is an example of how to define an HNSW vector index on a JSON based search index.

var jsonIndex = RediSearchIndex
    .OnJson()
    .ForKeysWithPrefix("test_hash_vector:")
    .WithSchema(
        s => s.Text("$.name", "name"),
        s => s.Vector("$.feature_embeddings", "feature_embeddings",
            VectorIndexAlgorithm.HNSW(
                type: VectorType.FLOAT32,
                dimensions: 512,
                distanceMetric: DistanceMetric.COSINE
            )
        )
    ).Build();

In the above example, the vector index is defined using only the required parameters. All of the optional index parameters are settable as optional method arguments.

Defining a FLAT index can be done by calling the FLAT builder method on the VectorIndexAlgorithm class when defining the vector field.

All of the relevant methods and parameters should have fully populated XMLDOC documentation.

Persisting a Vector

The following examples were pulled from the integration test suite.

HASH

_db.HashSet($"test_hash_vector:{vec.Name}", new[]
{
    new HashEntry("name", vec.Name),
    new HashEntry("feature_embeddings", vec.FileBytes)
});

JSON

_db.JsonSet($"test_json_vector:{vec.Name}", new
{
    name = vec.Name,
    feature_embeddings = vec.FileFloats
});

Executing a Query

I'm assuming that you already have a vector to query with as it's currently beyond the scope of this document to demonstrate the creation of feature embeddings.