Rag.NET.VectorStores.Redis 1.0.0

Rag.NET.VectorStores.Redis

Redis vector store for Rag.NET over RediSearch: dense cosine search on an HNSW index, using the Redis you already run rather than a second datastore.

Requires the RediSearch module — Redis Stack, or Redis 8 and later, where it is built in. Plain Redis answers FT.CREATE with an unknown-command error.

Install

dotnet add package Rag.NET.VectorStores.Redis

Install alongside the core pipeline package (dotnet add package Rag.NET), which supplies the AddRagNet(...) builder the store registers into.

Setup

Inside your AddRagNet(...) builder callback:

using Rag.NET.VectorStores.Redis;

rag.UseRedis(
    configuration:    "localhost:6379",
    indexName:        "ragnet-idx",
    vectorDimensions: 1536);

If Redis is already in the application — the case this store exists for — hand it the connection you have. The store does not dispose a multiplexer it did not create:

var redis = ConnectionMultiplexer.Connect("localhost:6379");
rag.UseRedis(redis, "ragnet-idx", 1536);

Metadata filtering needs keys declared up front

Every other vector store in this library accepts a MetadataFilter and just filters on whatever keys are in it. Redis cannot: RediSearch only filters on attributes its schema declares, and a hash field is not one unless the index was told about it at creation time. So UseRedis takes an extra, optional argument naming the metadata keys you intend to filter on:

rag.UseRedis(
    configuration:          "localhost:6379",
    indexName:              "ragnet-idx",
    vectorDimensions:       1536,
    filterableMetadataKeys: ["tenant"]);

Each declared key becomes a case-sensitive md_<key> TAG attribute on the index (e.g. tenant → md_tenant). Case-sensitive because metadata values are Base64Url-encoded before they are stored as tags, and Base64Url's alphabet uses both letter cases — a case-folding TAG field would match two genuinely different values whose tokens happen to be case-variants of one another.

Two consequences follow directly from keys being fixed at creation time:

• Filtering on an undeclared key throws InvalidOperationException naming the key, rather than silently returning an unfiltered page. MetadataFilter = { ["tenant"] = "acme" } against a store constructed without "tenant" in filterableMetadataKeys fails loudly, at query time.
• An index built before a key was declared fails InitializeAsync. Adding a key to filterableMetadataKeys does not retroactively add the TAG attribute to an existing index — InitializeAsync checks every declared key against the live schema and throws InvalidOperationException if one is missing. The index must be dropped and recreated (and its documents re-ingested) before that key can be filtered on.

This is Redis-only. Every other backend in this library either has a native document/map type it can filter against directly (PgVector, Weaviate, Chroma) or applies the filter as a metadata predicate outside a fixed schema (Qdrant, Pinecone, Azure AI Search); RediSearch is the one query engine here that refuses to match against an attribute its schema never named.

Example

Create the index once at startup. InitializeAsync is idempotent: an existing index is left alone, because re-creating it would discard every stored vector.

using Microsoft.Extensions.DependencyInjection;
using Rag.NET.Abstractions;
using Rag.NET.VectorStores.Redis;

var store = provider.GetRequiredService<IVectorStore>() as RedisVectorStore;
await store!.InitializeAsync();

var results = await pipeline.RetrieveAsync("open incidents", new RetrievalOptions
{
    TopK     = 5,
    MinScore = 0.6,
});

Scores are similarities, not distances

RediSearch returns vector_score as a cosine distance in [0, 2] — 0 is identical, larger is worse, the opposite direction from every score in this library. This store converts it to 1 - distance and reports ordinary cosine similarity, so MinScore means here what it means everywhere else.

Hybrid search is deliberately not offered: RediSearch's text scoring is TF-IDF-shaped rather than the BM25 the hybrid arm fuses, so the pipeline falls back to its own BM25 arm instead of fusing a score the store cannot describe.

Full guide

• Vector stores