{ "id": "dotnet", "title": "Redis semantic cache with NRedisStack", "url": "https://redis.io/docs/latest/develop/use-cases/semantic-cache/dotnet/", "summary": "Build a Redis-backed semantic cache for LLM responses in C# with NRedisStack and ONNX Runtime", "tags": [ "docs", "develop", "stack", "oss", "rs", "rc" ], "last_updated": "2026-06-01T09:32:08+01:00", "children": [], "page_type": "content", "content_hash": "2fca6053692cf37c97f8df6fee42c547e61005f4717601b997facf1466bccccf", "sections": [ { "id": "overview", "title": "Overview", "role": "overview", "text": "This guide shows you how to build a small Redis-backed semantic cache for LLM responses in C# / .NET with [`NRedisStack`](https://redis.io/docs/latest/develop/clients/dotnet/nredisstack) and [ONNX Runtime](https://onnxruntime.ai/) running the [`sentence-transformers/all-MiniLM-L6-v2`](https://huggingface.co/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." }, { "id": "overview", "title": "Overview", "role": "overview", "text": "Each cache entry is stored as a single Redis [Hash](https://redis.io/docs/latest/develop/data-types/hashes) at `cache:`. 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](https://redis.io/docs/latest/develop/ai/search-and-query) index covers the embedding field and every metadata field, so one [`FT.SEARCH`](https://redis.io/docs/latest/commands/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.\n\nThe lookup is thresholded: [`FT.SEARCH`](https://redis.io/docs/latest/commands/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.\n\nThe embedder is [ONNX Runtime](https://onnxruntime.ai/) loading the ONNX export of [`sentence-transformers/all-MiniLM-L6-v2`](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) (mirrored at [`Xenova/all-MiniLM-L6-v2`](https://huggingface.co/Xenova/all-MiniLM-L6-v2) on Hugging Face) paired with the `BertTokenizer` from [`Microsoft.ML.Tokenizers`](https://www.nuget.org/packages/Microsoft.ML.Tokenizers). This is the same 384-dimensional encoder the [Python example](https://redis.io/docs/latest/develop/use-cases/semantic-cache/redis-py), the [Node.js example](https://redis.io/docs/latest/develop/use-cases/semantic-cache/nodejs), and the [Java/Jedis example](https://redis.io/docs/latest/develop/use-cases/semantic-cache/java-jedis) 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.\n\nThat gives you:\n\n* A single round trip for lookup — vector KNN + metadata pre-filter in one [`FT.SEARCH`](https://redis.io/docs/latest/commands/ft.search).\n* 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.\n* Tenant, locale, and model-version isolation enforced inside the query, not in application code — a write under one tenant cannot be served to another.\n* Bounded memory: every entry has an [`EXPIRE`](https://redis.io/docs/latest/commands/expire) TTL, and a database-level [eviction policy](https://redis.io/docs/latest/develop/reference/eviction) (LRU / LFU) caps the cache size under pressure." }, { "id": "how-it-works", "title": "How it works", "role": "content", "text": "A query goes through three stages: **embed**, **lookup**, and (on a miss) **call the LLM and write back**." }, { "id": "hit-path-the-goal", "title": "Hit path (the goal)", "role": "content", "text": "1. The application calls `embedder.EncodeOne(prompt)` to turn the incoming text into a 384-element `float[]`.\n2. `cache.Lookup(queryVec, tenant, locale, modelVersion, \"ok\", threshold)` runs [`FT.SEARCH`](https://redis.io/docs/latest/commands/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.\n3. 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`](https://redis.io/docs/latest/commands/hincrby) on `hit_count` and an [`EXPIRE`](https://redis.io/docs/latest/commands/expire) refresh inside a [`MULTI/EXEC`](https://redis.io/docs/latest/commands/multi) (built with `IDatabase.CreateTransaction()`), so a frequently used answer keeps its TTL and the demo UI can see which entries are load-bearing.\n4. The LLM is not called at all. The application returns the cached response to the user." }, { "id": "miss-path", "title": "Miss path", "role": "content", "text": "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:\n\n1. Calls the LLM with the prompt.\n2. 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`](https://redis.io/docs/latest/commands/hset) and an [`EXPIRE`](https://redis.io/docs/latest/commands/expire) TTL inside a single [`MULTI/EXEC`](https://redis.io/docs/latest/commands/multi) so the entry never lands without a TTL on a partial failure.\n3. Returns the LLM's response to the user. The next semantically similar prompt under the same metadata scope will be a hit." }, { "id": "the-cache-helper", "title": "The cache helper", "role": "content", "text": "The `RedisSemanticCache` class wraps the Redis Search index and the lookup / write flow\n([source](https://github.com/redis/docs/blob/main/content/develop/use-cases/semantic-cache/dotnet/RedisSemanticCache.cs)):\n\n[code example]" }, { "id": "data-model", "title": "Data model", "role": "content", "text": "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.\n\n[code example]\n\nThe Redis Search index schema treats every field as queryable in its natural type:\n\n[code example]" }, { "id": "the-query", "title": "The query", "role": "content", "text": "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:\n\n[code example]\n\n`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." }, { "id": "the-mock-llm", "title": "The mock LLM", "role": "content", "text": "To make the latency and token savings visible without requiring an API key, `MockLLM.cs` provides a deterministic stand-in\n([source](https://github.com/redis/docs/blob/main/content/develop/use-cases/semantic-cache/dotnet/MockLLM.cs)):\n\n[code example]\n\nThe 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." }, { "id": "pre-seeding-the-cache", "title": "Pre-seeding the cache", "role": "content", "text": "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\n([source](https://github.com/redis/docs/blob/main/content/develop/use-cases/semantic-cache/dotnet/SeedCache.cs)):\n\n[code example]\n\nThe 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." }, { "id": "the-interactive-demo", "title": "The interactive demo", "role": "content", "text": "`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:\n\n* Type a prompt and toggle metadata: tenant, locale, model version. Each combination is a separate cache namespace inside the same index.\n* 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.\n* 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.\n* Watch the cumulative panel build up: total queries, cache hits, cache misses, hit ratio, tokens not spent, LLM milliseconds not waited.\n* Inspect every cached entry, including remaining TTL and total hit count, and drop individual entries to simulate eviction.\n\nThe 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 `` entry in the `.csproj`, so `dotnet bin/Release/net8.0/SemanticCacheDemo.dll` works from any working directory. Endpoints:\n\n| Endpoint | What it does |\n|-----------------|-------------------------------------------------------------------------------|\n| `GET /state` | Index info and the full list of cached entries. |\n| `POST /query` | Embed the prompt, run `FT.SEARCH`, on miss call the LLM and write back. |\n| `POST /reset` | Drop every cached entry and re-seed from the FAQ list. |\n| `POST /drop` | Delete a single cached entry by id. |" }, { "id": "run-the-demo-locally", "title": "Run the demo locally", "role": "content", "text": "1. Clone the [`redis/docs`](https://github.com/redis/docs) repository and change into the example\n directory:\n\n [code example]\n\n2. Make sure a Redis instance with the Redis Search module is running locally on\n port 6379. [Redis Stack](https://redis.io/docs/latest/operate/oss_and_stack/install/install-stack) or\n [Redis 8 with Search](https://redis.io/docs/latest/develop/ai/search-and-query) both work.\n\n3. Build the project. This pulls `NRedisStack`, `Microsoft.ML.OnnxRuntime`, and\n `Microsoft.ML.Tokenizers` from NuGet. The first build takes a couple of minutes:\n\n [code example]\n\n4. Run the demo. The first run downloads the ONNX export of\n `sentence-transformers/all-MiniLM-L6-v2` (~90 MB) and the matching BERT\n `vocab.txt` into a local `model_cache/` directory next to the binary; every\n subsequent run is offline:\n\n [code example]\n\n Or run the built binary directly:\n\n [code example]\n\n5. Open and try some queries:\n\n * **\"What is your return policy?\"** — exact match against the seed, distance ≈ 0,\n hit at any threshold.\n * **\"How fast is delivery?\"** — paraphrase of the shipping seed; distance\n around 0.30, hit at the default threshold of 0.5.\n * **\"How do I return an item?\"** — slightly looser paraphrase of the returns\n seed; distance around 0.49, still a hit at the default threshold. Slide\n the threshold down to 0.4 to see this one flip to a miss.\n * **\"What payment methods do you accept?\"** — unrelated to anything in the\n seed; distance > 0.6, so you'll see a miss, the mock LLM kicks in for\n ~1.5 s, the new answer is cached, and a follow-up of the same question\n is now an immediate hit.\n * Switch the **Tenant** dropdown to `globex` or `initech` and re-ask any\n seeded question — the result flips to a miss because the cache entries\n live under `acme`. That's the metadata pre-filter at work inside `FT.SEARCH`.\n\nThe 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." } ], "examples": [ { "id": "the-cache-helper-ex0", "language": "csharp", "code": "using NRedisStack.RedisStackCommands;\nusing StackExchange.Redis;\nusing SemanticCacheDemo;\n\nvar mux = ConnectionMultiplexer.Connect(\"localhost:6379\");\nvar db = mux.GetDatabase();\n\nvar embedder = await LocalEmbedder.CreateAsync(); // sentence-transformers/all-MiniLM-L6-v2\n\nvar cache = new RedisSemanticCache(\n db,\n indexName: \"semcache:idx\",\n keyPrefix: \"cache:\",\n vectorDim: 384,\n distanceThreshold: 0.5, // cosine distance, lower = stricter\n defaultTtlSeconds: 3600);\n\n// One-time index setup (idempotent).\ncache.CreateIndex();\n\n// 1) Embed the prompt.\nstring prompt = \"How do I return an item?\";\nfloat[] queryVec = embedder.EncodeOne(prompt);\n\n// 2) Look up under a metadata scope. The TAG filter and the KNN\n// travel together in one FT.SEARCH.\nvar result = cache.Lookup(\n queryVec, tenant: \"acme\", locale: \"en\",\n modelVersion: \"gpt-4.5-2026\");\n\nstring response;\nif (result is CacheHit hit)\n{\n response = hit.Response;\n Console.WriteLine($\"hit ({hit.Distance:F3}): {response}\");\n}\nelse\n{\n // 3a) Miss — call the LLM. (Use your real client here.)\n response = CallLlm(prompt);\n\n // 3b) Cache the new entry. Reuses the same embedding bytes the\n // lookup used, so we don't pay the encoder twice.\n cache.Put(\n prompt: prompt,\n response: response,\n embedding: queryVec,\n tenant: \"acme\",\n locale: \"en\",\n modelVersion: \"gpt-4.5-2026\");\n}", "section_id": "the-cache-helper" }, { "id": "data-model-ex0", "language": "text", "code": "cache:7c3f8a1b9e02\n prompt=How do I return an item?\n response=You can return any unworn item within 30 days...\n tenant=acme\n locale=en\n model_version=gpt-4.5-2026\n safety=ok\n created_ts=1715990400.123\n hit_count=4\n embedding=<384 × float32 little-endian bytes>", "section_id": "data-model" }, { "id": "data-model-ex1", "language": "text", "code": "FT.CREATE semcache:idx\n ON HASH PREFIX 1 cache:\n SCHEMA\n prompt TEXT\n response TEXT\n tenant TAG\n locale TAG\n model_version TAG\n safety TAG\n created_ts NUMERIC SORTABLE\n hit_count NUMERIC SORTABLE\n embedding VECTOR HNSW 6 TYPE FLOAT32 DIM 384 DISTANCE_METRIC COSINE", "section_id": "data-model" }, { "id": "the-query-ex0", "language": "csharp", "code": "// `.` and `-` (and other punctuation) are TAG-value syntax in Redis\n// Search and must be backslash-escaped — `gpt-4.5-2026` raw would\n// be parsed as three tokens. The helper's EscapeTagValue does this\n// for every TAG value; the literal below shows what the parser\n// actually sees on the wire.\nvar query = new Query(\n \"(@tenant:{acme} @locale:{en} @model_version:{gpt\\\\-4\\\\.5\\\\-2026} @safety:{ok})\"\n + \"=>[KNN 1 @embedding $vec AS distance]\")\n .ReturnFields(\n \"prompt\", \"response\", \"tenant\", \"locale\",\n \"model_version\", \"hit_count\", \"distance\")\n .SetSortBy(\"distance\", ascending: true)\n .Limit(0, 1)\n .AddParam(\"vec\", LocalEmbedder.ToBytes(queryVec))\n .Dialect(2);\n\nvar result = db.FT().Search(\"semcache:idx\", query);", "section_id": "the-query" }, { "id": "the-mock-llm-ex0", "language": "csharp", "code": "using SemanticCacheDemo;\n\nvar llm = new MockLLM(modelVersion: \"gpt-4.5-2026\", latencyMs: 1500.0);\nvar response = llm.Complete(\"What is your return policy?\");\n// response.Text — the templated answer text\n// response.LatencyMs — wall-clock time the call took\n// response.TotalTokens — estimated prompt + completion tokens", "section_id": "the-mock-llm" }, { "id": "pre-seeding-the-cache-ex0", "language": "csharp", "code": "cache.CreateIndex();\nSeedCache.Seed(cache, embedder, tenant: \"acme\", locale: \"en\");", "section_id": "pre-seeding-the-cache" }, { "id": "run-the-demo-locally-ex0", "language": "bash", "code": "git clone https://github.com/redis/docs.git\n cd docs/content/develop/use-cases/semantic-cache/dotnet", "section_id": "run-the-demo-locally" }, { "id": "run-the-demo-locally-ex1", "language": "bash", "code": "dotnet build -c Release", "section_id": "run-the-demo-locally" }, { "id": "run-the-demo-locally-ex2", "language": "bash", "code": "dotnet run -c Release", "section_id": "run-the-demo-locally" }, { "id": "run-the-demo-locally-ex3", "language": "bash", "code": "dotnet bin/Release/net8.0/SemanticCacheDemo.dll", "section_id": "run-the-demo-locally" } ] }