How to cache Haystack LLM calls with Crowkis
Add a semantic cache to Haystack so repeated and reworded questions are served for free, no rewrite, self-hosted.
Your users ask the same things all day, phrased a hundred different ways. If you build with Haystack, most of that repetition is invisible in your code but very visible on your bill. A semantic cache in front of your model calls fixes it.
The lowest-friction path is the OpenAI-compatible gateway: point Haystack's base URL at Crowkis and every model call flows through a semantic cache. Repeated and reworded prompts are served from cache with no upstream call; new ones pass through and get cached.
# point Haystack at the Crowkis gateway
base_url = "http://127.0.0.1:6380/v1" # semantic cache in front of your providerIn plain words. You don't restructure your Haystack app. You change where the calls go, and repeats stop costing money.
On repetitive workloads this cuts LLM costs up to 60-70% on repetitive workloads, and every hit comes back with a confidence score so reuse stays safe. It's one self-hosted binary, Redis-compatible, free to run.