← writing

Cut your LLM bill with semantic caching (video)

2026-07-12 · 4 min · 39 views

llmsemantic-cacheredisfinopsvideo

If your AI app answers the same question hundreds of times a day, you are paying the LLM hundreds of times for the same answer. Semantic caching fixes that, and I recorded a full walkthrough of how it works, end to end.

The walkthrough is in Brazilian Portuguese.

What semantic caching is

Before calling the LLM, you ask the cache one question: has anyone already asked something close enough to this? If yes, you return the stored answer in milliseconds and spend zero tokens. If no, you call the model and store the answer for next time. Search before, store after, two API calls.

The key word is semantic. "What is machine learning", "explain ML to me", and the same question in another language are three different strings but one meaning, so they should cost one generation, not three.

What the video walks through

The math

Savings are simple: output token cost times your cache hit rate. In FAQ, support, and assistant traffic, a real slice of questions is semantically repeated, so you stop paying to generate those answers again and again.

Large customers running semantic caching in production have seen 30%+ token savings. That is an observed result, never a guarantee. The part I like most: it scales with success. The more your app grows, the more repeated questions arrive, the more the cache hits.

Try it

Run it against your own traffic, measure the hit rate, and tell me what you see.