Redis recommendation engine with Lettuce

Build a Redis-backed recommendation engine in Java with Lettuce and DJL (HuggingFace tokenizers + ONNX Runtime)

This guide shows you how to build a small Redis-backed product recommendation service in Java with the Lettuce client library and the Deep Java Library (DJL) with its HuggingFace tokenizer integration and the ONNX Runtime inference engine. It includes a local web server built on the JDK's com.sun.net.httpserver 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 float32 vector. DJL's Predictor runs the HuggingFace tokenizer and ONNX Runtime inference end-to-end.
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, affinityWeight) 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, ewmaAlpha, affinityStep) 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):

import io.lettuce.core.RedisClient;
import io.lettuce.core.api.StatefulRedisConnection;
import io.lettuce.core.codec.ByteArrayCodec;
import io.lettuce.core.codec.RedisCodec;
import io.lettuce.core.codec.StringCodec;

// One client, two connections: a String/String one for structured
// commands and FT.* index management, plus a String/byte[] one for
// every command that touches the binary embedding field (including
// FT.SEARCH, whose $vec parameter is raw bytes too).
RedisClient client = RedisClient.create("redis://localhost:6379");
StatefulRedisConnection<String, String> conn = client.connect();
StatefulRedisConnection<String, byte[]> binConn = client.connect(
        RedisCodec.of(StringCodec.UTF8, ByteArrayCodec.INSTANCE));

Recommender recommender = new Recommender(conn, binConn,
        "recommend:idx", "product:", "user:", 384);

LocalEmbedder embedder = new LocalEmbedder();   // all-MiniLM-L6-v2 via DJL + ONNX Runtime

// One-time index setup (idempotent).
recommender.createIndex();

// Embed the natural-language query.
float[] 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
// (price range, rating), and TEXT (textMatch against textField) —
// Redis applies them all in front of the KNN.
Recommender.RetrieveOptions opts = new Recommender.RetrieveOptions();
opts.category = "outerwear";
opts.inStockOnly = true;
opts.minPrice = 50.0;
opts.maxPrice = 200.0;
opts.textMatch = "waterproof";   // TEXT pre-filter on @description
opts.k = 10;
List<Recommender.Candidate> candidates = recommender.candidateRetrieve(queryVec, opts);

// Record a click — updates the user's session vector and category
// affinity atomically; the next call to candidateRetrieve sees it.
recommender.recordClick("alice", "p001", 0.4, 1.0);

// Pull user features so the next retrieval can blend the session
// vector into the query and apply the category-affinity re-rank.
Recommender.UserFeatures features = recommender.getUserFeatures("alice");
opts.sessionVec = features.sessionVec;
opts.sessionWeight = 0.3;
candidates = recommender.candidateRetrieve(queryVec, opts);
candidates = recommender.rerank(candidates, features, 0.15);

// Hot embedding refresh — overwrite the vector for one product; the
// HNSW index reflects the change on the next FT.SEARCH.
float[] newVector = embedder.encodeOne("heavy-duty arctic expedition parka");
recommender.refreshEmbedding("p001", newVector);

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 x 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 x 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, Redis applies the filter first and then KNN-ranks only the matching documents.

FT.SEARCH recommend:idx
  "(@category:{outerwear} @in_stock:{true} @price:[50.0 200.0])
     =>[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 is the cosine distance (0 means identical, 2 means opposite), so the result is sorted ascending and the top row is the closest candidate to the query.

The Lettuce equivalent uses the typed ftSearch method on the binary-codec connection's RediSearchCommands. The query expression is passed as the value type (byte[] on the binary connection, so UTF-8-encoded) and the binary $vec parameter goes through SearchArgs.Builder.param(...):

String knn = filterClause + "=>[KNN " + k + " @embedding $vec AS vector_score]";
SearchArgs<String, byte[]> args = SearchArgs.<String, byte[]>builder()
        .param("vec", Recommender.floatsToBytes(queryVec))
        .returnField("name").returnField("description")
        .returnField("category").returnField("brand")
        .returnField("price").returnField("rating")
        .returnField("in_stock").returnField("vector_score")
        .sortBy(SortByArgs.<String>builder().attribute("vector_score").build())
        .limit(0, k)
        .dialect(QueryDialects.DIALECT2)
        .build();
SearchReply<String, byte[]> reply = binConn.sync()
        .ftSearch("recommend:idx",
                knn.getBytes(StandardCharsets.UTF_8), args);

Lettuce specifics: binary fields and pipelining

Two things in the helper change shape relative to the Jedis port:

1. Codec choice for the binary embedding field

Lettuce's default StringCodec UTF-8-decodes every hash value, which would corrupt the raw float32 bytes that the Redis Search vector field expects. Following the Lettuce vector-search reference, the helper opens a second connection bound to a <String, byte[]> codec (built with RedisCodec.of(StringCodec.UTF8, ByteArrayCodec.INSTANCE)) and routes every command that reads or writes the embedding field — including FT.SEARCH, whose $vec parameter is raw bytes too — through that connection. The structured fields share the same hash and are written through the same binary connection as their UTF-8 bytes so Redis sees an identical wire format to what the Jedis port writes.

RedisClient client = RedisClient.create("redis://localhost:6379");
StatefulRedisConnection<String, String> conn = client.connect();
StatefulRedisConnection<String, byte[]> binConn = client.connect(
        RedisCodec.of(StringCodec.UTF8, ByteArrayCodec.INSTANCE));

2. Pipelining with setAutoFlushCommands(false)

Jedis pipelines via a dedicated Pipeline object you obtain with client.pipelined(). Lettuce instead drives pipelining at the connection level: turn auto-flush off, queue async commands, flush, then await the futures.

RedisAsyncCommands<String, byte[]> async = binConn.async();
binConn.setAutoFlushCommands(false);
try {
    List<RedisFuture<Long>> futures = new ArrayList<>();
    for (Product p : products) {
        Map<String, byte[]> fields = ...;          // name, description, ..., embedding
        futures.add(async.hset(productKey(p.id), fields));
    }
    binConn.flushCommands();
    for (RedisFuture<Long> f : futures) f.get();
} finally {
    binConn.setAutoFlushCommands(true);
}

The toggle is connection-wide, so the helper owns its binary connection rather than borrowing a shared one — any other thread issuing commands on the same connection while auto-flush is off would be stalled until flushCommands() is called.

A note on FT.INFO

Lettuce 7's RediSearchCommands exposes typed wrappers for most FT.* commands, but FT.INFO isn't one of them. The helper dispatches the raw command via connection.sync().dispatch(...) with a NestedMultiOutput so it can parse the alternating-pair reply, the same approach the streaming Lettuce port uses for XAUTOCLAIM's extended reply.

The local embedder

The LocalEmbedder class wraps DJL's HuggingFace text-embedding pipeline so the rest of the helper can hand it a string and get back a unit-normalised float[] (source):

Criteria<String, float[]> criteria = Criteria.builder()
        .setTypes(String.class, float[].class)
        .optModelUrls("djl://ai.djl.huggingface.onnxruntime/sentence-transformers/all-MiniLM-L6-v2")
        .optEngine("OnnxRuntime")
        .optTranslatorFactory(new TextEmbeddingTranslatorFactory())
        .optProgress(new ProgressBar())
        .build();
ZooModel<String, float[]> model = criteria.loadModel();
Predictor<String, float[]> predictor = model.newPredictor();
float[] vector = predictor.predict("warm waterproof jacket for hiking");

The model URL routes through DJL's model zoo. The ai.djl.huggingface.onnxruntime group ID pulls a pre-converted ONNX bundle (model weights plus tokenizer.json) so we don't need a separate conversion step. optEngine("OnnxRuntime") pins the inference engine to ONNX Runtime, which means the demo only drags in the small ONNX Runtime native library rather than the much heavier PyTorch native library that ai.djl.huggingface.pytorch would require.

Predictor is not thread-safe; the wrapper guards calls with synchronized because the JDK HttpServer dispatches each request on a worker thread. For higher concurrency in production you'd hold a pool of Predictor instances backed by one ZooModel.

The catalogue builder

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

List<CatalogSeed.Seed> seeds = CatalogSeed.all();
List<String> texts = new ArrayList<>();
for (CatalogSeed.Seed s : seeds) {
    texts.add(CatalogSeed.embedTextFor(s));    // "<name>. <description>"
}
try (LocalEmbedder embedder = new LocalEmbedder()) {
    float[][] vectors = embedder.encodeMany(texts);
    Catalog.write(Path.of("catalog.json"),
            embedder.getModelName(), vectors[0].length, seeds, vectors);
}

In production the equivalent lives in an offline pipeline: embed once on catalogue 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 shared catalog.json wire format (model name, dim, list of products with base64-encoded float32 LE bytes for the vector) is identical to what the Python, Node, and Go ports produce, so you can re-use any port's catalog with the Lettuce demo as long as the embedding model matches.

The interactive demo

DemoServer runs com.sun.net.httpserver.HttpServer with a 16-thread executor and 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 RedisClient, two StatefulRedisConnections (regular + binary-codec), one LocalEmbedder, and one Recommender for the lifetime of the process. Endpoints:

Prerequisites

Before running the demo, make sure that:

• Redis 7.0 or later with the Redis Search module is running and accessible. By default the demo connects to localhost:6379. Redis Stack or Redis 8 with Search both work.
• JDK 17 or later is installed (the demo's inline HTML uses text blocks, which require JDK 15+; 17+ keeps the demo on a current LTS).
• Maven 3.9 or later for dependency resolution. DJL pulls in a couple of dozen transitive jars, so a manual javac -cp ... build is impractical; the pom.xml shipped next to the source files lets Maven handle that. Lettuce 7.x is required — the typed FT.* API used by this demo (ftCreate, ftSearch, ftDropindex, ftTagvals, the SearchArgs / FieldArgs / CreateArgs builders) lives in io.lettuce.core.search and was introduced in the 7.x release line.

If your Redis server is running elsewhere, start the demo with --redis-host and --redis-port.

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/java-lettuce
   

2. Build the project. The first build downloads the dependency graph (Lettuce 7, DJL API, the HuggingFace tokenizer extension, the ONNX Runtime engine, and Gson):

   mvn -DskipTests package
   

3. Generate the catalogue with pre-computed embeddings. The first run downloads the all-MiniLM-L6-v2 ONNX bundle (~80 MB) into the local DJL cache (~/.djl.ai/cache/):

   mvn -DskipTests exec:java -Dexec.mainClass=BuildCatalog
   

4. Start the demo server:

   mvn -DskipTests exec:java -Dexec.mainClass=DemoServer
   

   Override the defaults with -Dexec.args="--port 8090 --redis-host 127.0.0.1".

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, or --index-name / --key-prefix to point the demo at a different prefix entirely.