SentinelGateway Semantic Cache — Vector Embedding Deduplication

SentinelGateway caches LLM prompts by meaning, not just bytes. A two-stage pipeline checks an exact SHA-256 hash first, then a cosine-similarity vector search over text-embedding-3-small embeddings stored in Redis. Prompts scoring at or above the 0.92 similarity threshold are served from cache in under 50ms with zero tokens billed. Cache keys are tenant-isolated and embeddings are generated from PII-scrubbed text only.

🧠 SUB-50MS VECTOR CACHE — Embedding Deduplication

Stop paying for the same question twice.

Exact-match caching misses paraphrases. Sentinel embeds every prompt and serves semantically identical requests from cache — at zero token cost and under 50ms.

No credit card required · Drop-in OpenAI SDK compatible · Live in 60 seconds.

Live similarity engine

AUTO-REPLAY

Cached prompt · yesterday 14:02

"How do I reset a user's password?"

Incoming prompt · just now

"What's the way to reset a user password?"

cosine similarity 0.00
0.0 threshold ≥ 0.92 1.0
SEMANTIC CACHE HIT
0 tokens billed 24ms

How it works

A two-stage cache pipeline evaluated in-line — exact match first, vector similarity second, upstream last.

1

Ingest & Normalize

Messages are canonicalized and PII-scrubbed, then hashed (SHA-256) with your tenant ID and model. An exact hit returns instantly — 0ms, zero embedding cost.

2

In-Flight Vector Lookup

On an exact miss, the scrubbed prompt is embedded with text-embedding-3-small and compared against your tenant's cached vectors in Redis. Similarity at or above 0.92 returns the cached completion.

3

Delivery & Audit

Cache hits return with cache_type: "semantic" and full trace metadata. Misses flow upstream, then write back asynchronously — so the next paraphrase is free.

The FinOps math, upfront

Repeat and paraphrased prompts are the silent majority of LLM spend. Cache them once.

85%

Cost reduction

Typical savings on chat-heavy, repetitive workloads.

<25ms

Cache latency

Exact hits return in single-digit milliseconds.

$0.00

Per cache hit

Zero tokens billed — exact and semantic hits alike.

0.92

Cosine threshold

Tunable per tenant — tighten for precision, loosen for savings.

Your next 10,000 repeat prompts are free.

Sign up in 60 seconds. Semantic caching is on by default — no configuration required.

Explore the platform