featuresMay 20, 2026· 3 min read

Caching what the model saw: multimodal image-plus-text lookups

Vision queries are expensive and repetitive, the same product photo, the same screenshot, asked about again and again. Crowkis caches image-plus-text lookups so a repeated visual question is a hit.

Multimodal calls are among the priciest a model offers, and in production they repeat constantly: the same product image run through 'describe this,' the same screenshot asked 'what's the error,' the same chart queried 'summarize this.' A text-only cache can't see the image, so it misses every time. Crowkis caches the image-plus-text pair, so the second identical visual question is a hit.

In plain words: When the same image and question come in again, Crowkis returns the saved answer instead of re-running an expensive vision call, it caches the picture, not just the words.

CSET accepts an IMAGE argument alongside the query, and CGET (and CIMGGET) match on the combination, the image's content and the accompanying text together. So 'what's in this photo' over an identical image returns the cached answer instead of re-running an expensive vision pass, while a different image is correctly a different key.

what repeated traffic costs without crowkis

1
'how do refunds work?'
2
model call · $$ · 2s
3
'what's the refund window?'
4
model call · $$ · 2s
5
'refund timeline?'
6
model call · $$ · 2s
7
the same answer, three times

Every paraphrase is a fresh bill, unless the cache understands meaning.

The economics are the same argument the whole product makes, sharpened by price: vision tokens cost more than text tokens, so the savings per avoided call are larger. Anywhere the same images recur, catalogues, dashboards, document pipelines, support screenshots, the multimodal cache turns a repeated expensive call into a cheap lookup.

The bottom line

An LLM cache that goes blind the moment an image appears isn't a cache for how models are actually used now. Crowkis remembers what the model saw, not just what it was told, because the picture is increasingly part of the question.