Redis recommendation engine with NRedisStack

Build a Redis-backed recommendation engine in C# with NRedisStack and SmartComponents.LocalEmbeddings

This guide shows you how to build a small Redis-backed product recommendation service in C# with NRedisStack (the Redis Stack wrapper on top of StackExchange.Redis) and the SmartComponents.LocalEmbeddings library, which ships a bundled 384-dimensional BAAI bge-micro-v2 ONNX model (no API key, no shared library to install). The other ports in this guide series use sentence-transformers/all-MiniLM-L6-v2, also a 384-dim encoder; the index schema is identical, so the same FT.SEARCH query shape works for both. It includes a local web server built on System.Net.HttpListener so you can embed a natural-language query, run a KNN retrieval with structured pre-filters in one round trip, feed clicks back as a session signal, and watch the next recommendation incorporate them immediately.

Overview

Each product is stored as a single Redis Hash at product:<id>. The hash holds the structured metadata (name, description, category, brand, price, rating, in-stock flag) alongside the raw float32 bytes of a 384-dimensional embedding. A single Redis Search index covers every field, so one FT.SEARCH call with a KNN clause does the vector similarity and the TAG / NUMERIC / TEXT pre-filtering in the same pass — no cross-store joins.

Per-user state lives in user:<id>:features: a session vector written as an exponentially weighted average of recently-clicked item embeddings, plus per-category affinity counters incremented atomically with HINCRBYFLOAT. FT.SEARCH does not read that hash directly; instead, the application reads it on the next request and passes the session vector to FT.SEARCH as the query parameter. The two-step is what lets a click feed the very next recommendation without a batch cycle or cache invalidation.

That gives you:

• A single round trip for retrieval — vector KNN + structured filters in one FT.SEARCH.
• Sub-millisecond hot path once the query is embedded; embedding the query is the bottleneck, and that's a model-side cost, not a Redis one.
• Real-time session signals — a click writes a new session vector and bumps an affinity counter; the next query reads them and folds them in.
• No-downtime embedding refresh — HSET on the vector field, and the HNSW index reflects the change on the next query.

How it works

There are two distinct paths: a query path runs every time the application wants a recommendation, and a click path runs every time the user interacts with a product.

Query path (per recommendation request)

1. The application calls embedder.EncodeOne(queryText) to turn a natural-language query into a 384-dimensional float[].
2. The application reads the user's session vector and affinities from the user features hash. If a session vector exists, it gets blended into the query vector with a tunable weight, so the user's recent clicks pull retrieval toward what they've been engaging with.
3. recommender.CandidateRetrieve(queryVec, opts) runs FT.SEARCH with a pre-filter clause built from the request's TAG / NUMERIC / TEXT inputs, followed by a KNN k @embedding $vec clause. Redis returns up to k candidates with the cosine distance to the query (lower is closer).
4. recommender.Rerank(candidates, userFeatures, weight) subtracts a log-scaled per-category affinity bonus from each candidate's distance and re-sorts the list closest-first. The log scaling keeps repeated clicks from running away with the ranking.

Click path (per user interaction)

When the user clicks a product, recommender.RecordClick(userId, productId) does the following:

1. Reads the clicked item's embedding from its hash.
2. Reads the user's previous session vector from the user features hash, blends the new click in via an exponentially weighted moving average, and writes the new session vector back with HSET. This is a read-modify-write — atomic against any single write but not against a concurrent click for the same user; in practice, per-user click streams don't generate the contention to make this matter, and if a deployment does, the read and write can be wrapped in WATCH/MULTI/EXEC or a small Lua script.
3. Bumps the per-category affinity counter with HINCRBYFLOAT — atomic, no read needed — and the click count with HINCRBY.

The next query path picks both changes up the next time it reads the user features hash.

Refreshing an item's embedding follows a similar shape: encode the new text, write the vector bytes back with HSET, and the HNSW index reflects the change on the next query without a rebuild.

The recommender helper

The Recommender class wraps the Redis Search index and the retrieval flow (source):

using StackExchange.Redis;
using RecommendationDemo;

var mux = ConnectionMultiplexer.Connect("localhost:6379");
var recommender = new Recommender(
    mux.GetDatabase(),
    indexName: "recommend:idx",
    keyPrefix: "product:");
recommender.CreateIndex();                 // idempotent

// SmartComponents.LocalEmbeddings ships its own 384-dim ONNX model
// (BAAI's bge-micro-v2) which the Embedder loads on construction. The
// other ports in this guide series use sentence-transformers/all-MiniLM-L6-v2,
// also 384-dim. Both encode to the same vector field; vectors aren't
// interchangeable across the two models, but each port produces its own
// catalog.json so cross-port mixing isn't an issue.
using var embedder = new Embedder();       // bge-micro-v2, 384-dim
var queryVec = embedder.EncodeOne("warm waterproof jacket for hiking");

// Retrieval: KNN with structured pre-filter in one round trip.
// Filters combine TAG (Category, Brand, InStockOnly), NUMERIC
// (MinPrice/MaxPrice, MinRating), and TEXT (TextMatch against
// TextField) — Redis applies them all in front of the KNN.
var candidates = recommender.CandidateRetrieve(queryVec, new RetrieveOptions
{
    Filter = new FilterOptions
    {
        Category    = "outerwear",
        InStockOnly = true,
        MinPrice    = 50,
        MaxPrice    = 200,
        TextMatch   = "waterproof",   // TEXT pre-filter on @description
    },
    K = 10,
});

// Record a click — updates the user's session vector and category
// affinity atomically; the next call to CandidateRetrieve sees it.
recommender.RecordClick("alice", "p001");

// Pull user features so the next retrieval can blend the session
// vector into the query and apply the category-affinity re-rank.
var features = recommender.GetUserFeatures("alice");
candidates = recommender.CandidateRetrieve(queryVec, new RetrieveOptions
{
    Filter = new FilterOptions { Category = "outerwear", InStockOnly = true },
    K = 10,
    SessionVec    = features.SessionVec,
    SessionWeight = 0.3,
});
candidates = recommender.Rerank(candidates, features, affinityWeight: 0.15);

// Hot embedding refresh — overwrite the vector for one product;
// the HNSW index reflects the change on the next FT.SEARCH.
var newVec = embedder.EncodeOne("heavy-duty arctic expedition parka");
recommender.RefreshEmbedding("p001", newVec);

Data model

Each product is one Redis Hash. The vector field is raw little-endian float32 bytes — no JSON wrapping — because the Redis Search vector encoding expects exactly that.

product:p001
  name=Alpine down parka
  description=Heavyweight 800-fill goose down parka...
  category=outerwear
  brand=northpeak
  price=289.0
  rating=4.7
  in_stock=true
  embedding=<384 × float32 little-endian bytes>

The Redis Search index schema treats every field as queryable in its natural type:

FT.CREATE recommend:idx
  ON HASH PREFIX 1 product:
  SCHEMA
    name        TEXT
    description TEXT
    category    TAG
    brand       TAG
    in_stock    TAG
    price       NUMERIC SORTABLE
    rating      NUMERIC SORTABLE
    embedding   VECTOR HNSW 6 TYPE FLOAT32 DIM 384 DISTANCE_METRIC COSINE

Per-user state is a separate hash. The session vector is stored as raw float32 bytes the same way; affinity counters are stored as plain numeric strings, one field per category, prefixed with aff: so they don't collide with anything else.

user:alice:features
  session_vec=<384 × float32 little-endian bytes>
  aff:outerwear=2.0
  aff:footwear=1.0
  last_clicked_id=p015
  last_clicked_category=footwear
  clicks=3

The query

The KNN clause is a hybrid query: a pre-filter expression in parentheses, then =>[KNN k @embedding $vec]. With dialect 2 (the default in NRedisStack 1.0+), Redis applies the filter first and then KNN-ranks only the matching documents.

FT.SEARCH recommend:idx
  "(@category:{outerwear} @in_stock:{true} @price:[50 200])
     =>[KNN 10 @embedding $vec AS vector_score]"
  PARAMS 2 vec <384-float32-bytes>
  SORTBY vector_score
  RETURN 8 name description category brand price rating in_stock vector_score
  DIALECT 2

When there's no filter, the pre-filter clause is just (*). vector_score returned by Redis is the cosine distance (0 means identical, 2 means opposite), so the result is sorted ascending — a score of 0.0 is a perfect match.

Binary fields with NRedisStack and StackExchange.Redis

The embedding field is binary, while everything else in the same hash is text. StackExchange.Redis handles mixed types in a single HSET automatically — pass byte[] for binary fields and string for text fields in the same HashEntry array:

using StackExchange.Redis;

db.HashSet("product:p001", new HashEntry[]
{
    new("name",      "Alpine down parka"),
    new("price",     "289.0"),
    new("embedding", Embedder.FloatsToBytes(vector)),   // byte[]
});

FT.SEARCH parameter values follow the same convention — pass the raw byte[] to Query.AddParam and NRedisStack includes it in the command body without any encoding.

The Recommender class uses IDatabase.CreateBatch() to pipeline the HSETs for the catalogue, so the initial seed of 37 products goes out in a single round trip.

The catalog builder

Item vectors are pre-computed once and stored in catalog.json so the demo server can boot quickly. BuildCatalog.cs and the build-catalog sub-command are the reference for how to do that — and are the code you'd adapt for a real catalog ingestion pipeline (source):

using var embedder = new Embedder();
var texts = CatalogSeed.All.Select(p => p.EmbedText()).ToList();
var vectors = embedder.EncodeMany(texts);

var file = new CatalogFile
{
    Model = embedder.ModelName,
    Dim   = embedder.Dim,
    Products = CatalogSeed.All.Select((p, i) => new CatalogProduct
    {
        Id = p.Id, Name = p.Name, Description = p.Description,
        Category = p.Category, Brand = p.Brand,
        Price = p.Price, InStock = p.InStock, Rating = p.Rating,
        EmbeddingB64 = Convert.ToBase64String(
            Embedder.FloatsToBytes(vectors[i])),
    }).ToList(),
};
File.WriteAllText("catalog.json", JsonSerializer.Serialize(file, opts));

In production the equivalent of this script lives in an offline pipeline: embed once on catalog updates and ship the vectors into Redis with HSET. The serving tier still embeds the query on each request, but that's usually fronted by a dedicated model server or batched at the API gateway rather than co-located with the data tier as it is in this demo.

The interactive demo

DemoServer.cs runs an HttpListener server with one demo user (demo). The HTML page lets you:

• Type a natural-language query and toggle filters: TAG (category, brand, in-stock), NUMERIC (price range, rating), and TEXT (the Description contains field, a phrase pre-filter against the description text index).
• Toggle session blending and category-affinity re-ranking independently to see what each layer contributes.
• Click any product card to record a click into the session. The page re-renders the user features panel immediately — the click wrote to the user features hash, and the next search reads that hash to fold the update in.
• Refresh a single product's embedding with new text and watch the ranking change on the next query.

The server holds one Embedder instance and one Recommender for the lifetime of the process. Endpoints:

Prerequisites

• Redis 7.4 or later with the Redis Search module (Redis Stack or Redis Community Edition).
• .NET 8 SDK or later.
• NuGet packages, all declared in RecommendationDemo.csproj: StackExchange.Redis, NRedisStack (the Redis Stack wrapper that exposes FT.* and JSON.* over the StackExchange.Redis multiplexer), and SmartComponents.LocalEmbeddings (in-process ONNX runtime that ships a 384-dimensional sentence encoder with no external download or API key).

Run the demo locally

1. Clone the redis/docs repository and change into the example directory:

   git clone https://github.com/redis/docs.git
   cd docs/content/develop/use-cases/recommendation-engine/dotnet
   

2. Make sure a Redis instance with the Redis Search module is running locally on port 6379. Redis Stack or Redis 8 with Search both work.

3. Generate the catalog with pre-computed embeddings. The first run loads the ONNX model bundled with SmartComponents.LocalEmbeddings (no external download):

   dotnet run -- build-catalog
   

4. Start the demo server:

   dotnet run -- --port 8084
   

5. Open http://localhost:8084 and try some queries:

   • "insulated down jacket for cold weather" — filtered to outerwear, in-stock only.
   • "comfortable shoes for trail running" — filtered to footwear.
   • Add Description contains: waterproof to either query above to see a TEXT pre-filter combine with the KNN.
   • Click a couple of products to seed a session, then re-run the same query with Blend session vector into query on and watch the ranking shift.
   • Use Refresh embedding to change a product's vector — for example, change the Alpine down parka's text to "heavy duty arctic expedition parka with hood" and re-run the jacket query to see the result move.

The server is read/write against your local Redis. The default index name is recommend:idx and product keys live under product:. Pass --no-reset to keep an existing index across restarts, --redis-host / --redis-port to point at a different Redis, or --index-name and --key-prefix to share a Redis instance with other ports without collisions.