Redis semantic cache with NRedisStack

Build a Redis-backed semantic cache for LLM responses in C# with NRedisStack and ONNX Runtime

This guide shows you how to build a small Redis-backed semantic cache for LLM responses in C# / .NET with NRedisStack and ONNX Runtime running the sentence-transformers/all-MiniLM-L6-v2 encoder locally. It includes a local web server built with the BCL's standard System.Net.HttpListener so you can send paraphrased prompts at a mock LLM, watch the cache decide hit or miss, sweep the cosine-distance threshold, and see the cumulative latency and token savings build up.

Overview

Each cache entry is stored as a single Redis Hash at cache:<id>. The hash holds the original prompt, the LLM's response, the raw float32 bytes of a 384-dimensional embedding of the prompt, and metadata fields — tenant, locale, model version, safety flag — plus a created_ts and a hit_count. A single Redis Search index covers the embedding field and every metadata field, so one FT.SEARCH call with a KNN clause does the vector lookup and the TAG pre-filter in the same round trip — no cross-store joins.

The lookup is thresholded: FT.SEARCH always returns the nearest entry that satisfies the filters, but the application only serves it as a hit when the reported cosine distance is at or below DistanceThreshold. Anything further away is treated as a miss; the caller runs the LLM and writes the new prompt, response, and embedding back to the same key pattern with a TTL.

The embedder is ONNX Runtime loading the ONNX export of sentence-transformers/all-MiniLM-L6-v2 (mirrored at Xenova/all-MiniLM-L6-v2 on Hugging Face) paired with the BertTokenizer from Microsoft.ML.Tokenizers. This is the same 384-dimensional encoder the Python example, the Node.js example, and the Java/Jedis example use. Embeddings produced by the implementations are semantically equivalent — paraphrase distances differ only at the second or third decimal place — so a cache populated by one demo can be queried by another against the same Redis instance.

That gives you:

• A single round trip for lookup — vector KNN + metadata pre-filter in one FT.SEARCH.
• Tens of milliseconds on a hit vs. a multi-second LLM call on a miss; the embedding step is the bottleneck either way, and that's a model-side cost, not a Redis one.
• Tenant, locale, and model-version isolation enforced inside the query, not in application code — a write under one tenant cannot be served to another.
• Bounded memory: every entry has an EXPIRE TTL, and a database-level eviction policy (LRU / LFU) caps the cache size under pressure.

How it works

A query goes through three stages: embed, lookup, and (on a miss) call the LLM and write back.

Hit path (the goal)

1. The application calls embedder.EncodeOne(prompt) to turn the incoming text into a 384-element float[].
2. cache.Lookup(queryVec, tenant, locale, modelVersion, "ok", threshold) runs FT.SEARCH with a TAG pre-filter and a KNN 1 clause. Redis returns the closest cached prompt that satisfies the filters along with its cosine distance.
3. If the distance is at or below the threshold, the cache returns a CacheHit record containing the cached response. The helper also runs an HINCRBY on hit_count and an EXPIRE refresh inside a MULTI/EXEC (built with IDatabase.CreateTransaction()), so a frequently used answer keeps its TTL and the demo UI can see which entries are load-bearing.
4. The LLM is not called at all. The application returns the cached response to the user.

Miss path

When the distance is above the threshold — or there is no candidate in scope at all — the helper returns a CacheMiss record instead, carrying the distance of the nearest candidate (if any) for logging. The application then:

1. Calls the LLM with the prompt.
2. Calls cache.Put(prompt, response, embedding, tenant, locale, modelVersion, ...). The same embedding the lookup used is reused — no re-encode. The helper writes the Hash with HSET and an EXPIRE TTL inside a single MULTI/EXEC so the entry never lands without a TTL on a partial failure.
3. Returns the LLM's response to the user. The next semantically similar prompt under the same metadata scope will be a hit.

The cache helper

The RedisSemanticCache class wraps the Redis Search index and the lookup / write flow (source):

using NRedisStack.RedisStackCommands;
using StackExchange.Redis;
using SemanticCacheDemo;

var mux = ConnectionMultiplexer.Connect("localhost:6379");
var db = mux.GetDatabase();

var embedder = await LocalEmbedder.CreateAsync();   // sentence-transformers/all-MiniLM-L6-v2

var cache = new RedisSemanticCache(
    db,
    indexName: "semcache:idx",
    keyPrefix: "cache:",
    vectorDim: 384,
    distanceThreshold: 0.5,    // cosine distance, lower = stricter
    defaultTtlSeconds: 3600);

// One-time index setup (idempotent).
cache.CreateIndex();

// 1) Embed the prompt.
string prompt = "How do I return an item?";
float[] queryVec = embedder.EncodeOne(prompt);

// 2) Look up under a metadata scope. The TAG filter and the KNN
//    travel together in one FT.SEARCH.
var result = cache.Lookup(
    queryVec, tenant: "acme", locale: "en",
    modelVersion: "gpt-4.5-2026");

string response;
if (result is CacheHit hit)
{
    response = hit.Response;
    Console.WriteLine($"hit ({hit.Distance:F3}): {response}");
}
else
{
    // 3a) Miss — call the LLM. (Use your real client here.)
    response = CallLlm(prompt);

    // 3b) Cache the new entry. Reuses the same embedding bytes the
    //     lookup used, so we don't pay the encoder twice.
    cache.Put(
        prompt: prompt,
        response: response,
        embedding: queryVec,
        tenant: "acme",
        locale: "en",
        modelVersion: "gpt-4.5-2026");
}

Data model

Each cache entry 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. The helper packs the float[] with BinaryPrimitives.WriteSingleLittleEndian so the byte order is pinned to little-endian regardless of host architecture; every supported .NET runtime today is little-endian, and the explicit encoding matches the bytes the Python, Node, Go, and Java ports write.

cache:7c3f8a1b9e02
  prompt=How do I return an item?
  response=You can return any unworn item within 30 days...
  tenant=acme
  locale=en
  model_version=gpt-4.5-2026
  safety=ok
  created_ts=1715990400.123
  hit_count=4
  embedding=<384 × float32 little-endian bytes>

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

FT.CREATE semcache:idx
  ON HASH PREFIX 1 cache:
  SCHEMA
    prompt         TEXT
    response       TEXT
    tenant         TAG
    locale         TAG
    model_version  TAG
    safety         TAG
    created_ts     NUMERIC SORTABLE
    hit_count      NUMERIC SORTABLE
    embedding      VECTOR HNSW 6 TYPE FLOAT32 DIM 384 DISTANCE_METRIC COSINE

The query

The lookup is a hybrid query: a TAG pre-filter expression in parentheses, then =>[KNN 1 @embedding $vec]. With DIALECT 2, Redis applies the filter first and KNN-ranks only the matching documents. In NRedisStack:

// `.` and `-` (and other punctuation) are TAG-value syntax in Redis
// Search and must be backslash-escaped — `gpt-4.5-2026` raw would
// be parsed as three tokens. The helper's EscapeTagValue does this
// for every TAG value; the literal below shows what the parser
// actually sees on the wire.
var query = new Query(
        "(@tenant:{acme} @locale:{en} @model_version:{gpt\\-4\\.5\\-2026} @safety:{ok})"
        + "=>[KNN 1 @embedding $vec AS distance]")
    .ReturnFields(
        "prompt", "response", "tenant", "locale",
        "model_version", "hit_count", "distance")
    .SetSortBy("distance", ascending: true)
    .Limit(0, 1)
    .AddParam("vec", LocalEmbedder.ToBytes(queryVec))
    .Dialect(2);

var result = db.FT().Search("semcache:idx", query);

distance is the cosine distance (0 means identical, 2 means opposite). The result is sorted ascending, so the top row is the closest candidate. The application inspects distance against the threshold and decides hit or miss in user code — Redis returns the row either way, and treating it as a hit or a miss is a policy decision the cache helper owns, not a server-side filter.

The mock LLM

To make the latency and token savings visible without requiring an API key, MockLLM.cs provides a deterministic stand-in (source):

using SemanticCacheDemo;

var llm = new MockLLM(modelVersion: "gpt-4.5-2026", latencyMs: 1500.0);
var response = llm.Complete("What is your return policy?");
// response.Text         — the templated answer text
// response.LatencyMs    — wall-clock time the call took
// response.TotalTokens  — estimated prompt + completion tokens

The mock sleeps for the configured latency, then keyword-matches against a small FAQ table to produce an answer. The deliberate slowness is what makes a hit visibly cheaper than a miss in the demo. In production code, you would replace MockLLM with your real client of choice — an HTTP call to OpenAI, Anthropic, an Azure OpenAI deployment, a self-hosted vLLM endpoint, anything — without changing the cache helper.

Pre-seeding the cache

In a real deployment the cache fills up organically: a first-time question is a miss, the LLM answers, and the response is written back. For the demo, SeedCache.cs pre-loads a small set of canonical FAQ prompts so the very first query lands on a hit (source):

cache.CreateIndex();
SeedCache.Seed(cache, embedder, tenant: "acme", locale: "en");

The seed list stores the canonical phrasing of each question ("What is your return policy?"). Paraphrases of any of these prompts ("How do I return an item?", "Can I get a refund?") embed close to the canonical entry, so the cache lookup serves the stored response without ever calling the model.

The interactive demo

Program.cs runs an HTTP server built on the BCL's System.Net.HttpListener — no ASP.NET Core, no Kestrel, no Minimal API. The HTML page lets you:

• Type a prompt and toggle metadata: tenant, locale, model version. Each combination is a separate cache namespace inside the same index.
• Slide the cosine-distance threshold and see hits flip to misses (and back) on the same prompt, with the actual distance reported on each query.
• Submit with Ask to run the full hit-or-miss path (calls the LLM on a miss, writes the answer back). Submit with Lookup only (no LLM) to sweep the threshold against a fixed prompt without polluting the cache.
• Watch the cumulative panel build up: total queries, cache hits, cache misses, hit ratio, tokens not spent, LLM milliseconds not waited.
• Inspect every cached entry, including remaining TTL and total hit count, and drop individual entries to simulate eviction.

The server holds one LocalEmbedder, one RedisSemanticCache, and one MockLLM for the lifetime of the process. The HTML page is shared with the Python, Node.js, Go, and Java demos; index.html ships next to the binary via a <None CopyToOutputDirectory> entry in the .csproj, so dotnet bin/Release/net8.0/SemanticCacheDemo.dll works from any working directory. Endpoints:

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/semantic-cache/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. Build the project. This pulls NRedisStack, Microsoft.ML.OnnxRuntime, and Microsoft.ML.Tokenizers from NuGet. The first build takes a couple of minutes:

   dotnet build -c Release
   

4. Run the demo. The first run downloads the ONNX export of sentence-transformers/all-MiniLM-L6-v2 (~90 MB) and the matching BERT vocab.txt into a local model_cache/ directory next to the binary; every subsequent run is offline:

   dotnet run -c Release
   

   Or run the built binary directly:

   dotnet bin/Release/net8.0/SemanticCacheDemo.dll
   

5. Open http://localhost:8092 and try some queries:

   • "What is your return policy?" — exact match against the seed, distance ≈ 0, hit at any threshold.
   • "How fast is delivery?" — paraphrase of the shipping seed; distance around 0.30, hit at the default threshold of 0.5.
   • "How do I return an item?" — slightly looser paraphrase of the returns seed; distance around 0.49, still a hit at the default threshold. Slide the threshold down to 0.4 to see this one flip to a miss.
   • "What payment methods do you accept?" — unrelated to anything in the seed; distance > 0.6, so you'll see a miss, the mock LLM kicks in for ~1.5 s, the new answer is cached, and a follow-up of the same question is now an immediate hit.
   • Switch the Tenant dropdown to globex or initech and re-ask any seeded question — the result flips to a miss because the cache entries live under acme. That's the metadata pre-filter at work inside FT.SEARCH.

The server is read/write against your local Redis. The default index name is semcache:idx and entry keys live under cache:. Flags mirror the Python, Node, Go, and Java demos: --no-reset to keep an existing cache across restarts, --threshold to change the default cosine-distance cutoff, --llm-latency-ms to make the mock LLM faster or slower for the demo, or --port to listen on a different port.