Adaptive thresholds: the cache tunes its own reuse bar over time
A fixed similarity threshold is wrong the day after you set it. Crowkis uses a three-tier scheme — per-intent base, complexity adjustment, and an EMA feedback loop — that learns the right bar and persists it.
Pick a similarity threshold and you've made a guess that ages badly: too loose and the cache serves near-misses, too tight and it misses safe reuse, and the right value drifts as your traffic shifts. Crowkis doesn't ask you to pick one number; it runs a three-tier scheme that converges on the right bar and keeps it current.
Tier one is a table of base thresholds, one per intent class, because a factual lookup and a creative request have different safe bars to begin with. Tier two adjusts for query complexity, nudging the bar up or down by up to 0.08 based on how hard the question is. Tier three is the learner: an exponential moving average over feedback, decaying old signal as new arrives, so the threshold tracks reality instead of a launch-day assumption.
Reuse only when meaning, structure, confidence, and trust all agree.
Crucially, what it learns persists. The tuned thresholds survive restarts, so the cache doesn't forget months of calibration because a pod rescheduled — it comes back as smart as it went down. The adaptation is per intent, so the system learns a different lesson for factual traffic than for personal traffic, which is the only granularity that makes the learning safe.
The bottom line
A threshold you set once is a guess with a timestamp. A threshold that learns is a system that gets more correct the longer it runs — and persists what it learned, so the calibration compounds instead of resetting.