//

NOTES

blog

//

NOTES

blog

Engineering

8 min

Caching evals without lying to yourself

Caching evals without lying to yourself

A cache hit is only safe if the prompt, the case and the model settings all match. How we hash all three, and the bug that taught us to.

A cache hit is only safe if the prompt, the case and the model settings all match. How we hash all three, and the bug that taught us to.

Mira Okafor

Skipping unchanged cases is the biggest speed-up we have. It is also the easiest way to ship a regression, if the cache key is wrong.

We hash the prompt template, the case and every model setting. Change the temperature and every case reruns.

tessel

Evals for LLM apps, on every commit.

Release notes, once a month

All systems normal · © 2026 Tessel Labs

tessel

tessel

Evals for LLM apps, on every commit.

Release notes, once a month

All systems normal · © 2026 Tessel Labs

tessel

Create a free website with Framer, the website builder loved by startups, designers and agencies.