A
Aesthetic AI
← All Articles • AI & Optimization • 5 min read

Zero-Cost Semantic Caching: Slashing LLM Bills Without Stale Answers

By AI Infrastructure Engineering

The Inefficiency of Repetitive LLM Generations

In customer support for aesthetic clinics and boutique hotels, inquiries follow a Pareto distribution:
  • "Where can I park my car?"
  • - "What time is checkout?"
  • "Can I exercise after Botox injections?"
  • Dispatching every identical or closely paraphrased question to an external LLM costs unnecessary money and introduces 2 to 4 seconds of generation latency.

    Two-Tier Semantic & Exact Caching

    Aesthetic AI introduces a dual-layer caching pipeline: 1. **Exact Canonical Query Hash**: Normalizes user input (lowercasing, punctuation stripping, language canonicalization) and checks Redis in <2ms. 2. **Cosine Vector Semantic Cache**: Generates a high-speed embedding of the query. If the cosine similarity to a recently answered question exceeds 0.94 within the same tenant context, the verified answer is returned immediately.

    Zero Credit Charge Policy

    Because cache hits incur negligible compute overhead on Cloudflare Workers, **tenants are charged 0 credits for cache hits**. This directly rewards clinics and hotels for having well-defined, frequently accessed FAQs.

    Instant Versioned Invalidation

    When a tenant updates their clinic pricing or hotel check-in policy, how do we prevent stale cached responses? Every cache entry is tagged with the tenant's current `knowledge_version`. As soon as a clinic coordinator publishes an updated price list, the `knowledge_version` increments atomically, invalidating all older cached answers instantly.