Twelve intents: why the cache treats a poem differently from a fact
One similarity threshold for all traffic is how caches embarrass themselves. Crowkis classifies every query into one of twelve intents, each with its own rules of reuse.
The single-threshold semantic cache has a structural flaw: 0.85 similarity means entirely different things for different queries. For 'capital of France?' it's a safe hit; for 'write my wedding toast' it's a plagiarism machine; for 'should I take this medication with alcohol?' it's a liability. Reuse safety isn't a number — it's a number per kind of question.
Crowkis classifies every query into one of twelve intent classes — factual, conceptual, transactional, creative, personal, temporal, and friends — before any matching happens. Each class carries its own base threshold, freshness expectations, and reuse posture: factual content reuses aggressively at high confidence; creative and personal intents face strict bars or no-reuse rules; temporal queries inherit tight TTLs by default.
Reuse only when meaning, structure, confidence, and trust all agree.
Classification feeds everything downstream: confidence gates read the intent's threshold, adaptive tuning adjusts per class rather than globally, eviction values entries by class economics, and the dashboard breaks down hits by intent so you can see where the cache earns. It's the routing layer of the whole intelligence stack.
The bottom line
The classifier itself is fast, deterministic, rule-grounded, and clamped — boring on purpose, because everything trusts it. Twelve buckets sounds simple. Most production incidents in naive caches trace back to not having them.