featuresMay 17, 2026· 3 min read

Confidence scoring: every hit arrives with a number you can gate on

A cache that only says 'hit' or 'miss' makes you trust it blindly. Crowkis returns a confidence score per hit, a geometric mean of five signals, so you decide the bar reuse must clear.

Most caches answer a binary question: is it a hit? That forces blind trust, you take the cached answer or you don't, with no sense of how safe the reuse actually is. Crowkis answers a richer question, returning a confidence score with every hit so 'should I trust this?' has a number behind it.

In plain words: Every cached answer comes with a score for how safe it is to reuse, built from five signals where any weak one tanks the total, so you set the bar and the cache respects it.

The score is a geometric mean of five signals, similarity, freshness, trust, validation, and domain-accuracy, and the geometric mean is the deliberate choice: it's unforgiving, because any single signal near zero drags the whole score down. A perfectly similar but stale answer, or a fresh but low-trust one, scores low, exactly as it should. One strong signal can't paper over a weak one.

the crowkis read path, five gates, every one can veto

1
incoming query
2
intent classifier
3
template match
4
HNSW neighbours
5
confidence gate
6
trust + freshness
7
answer · <1ms
8
(nil) → your model

Reuse only when meaning, structure, confidence, and trust all agree.

You consume it with WITHCONFIDENCE and gate on it per intent: serve factual hits above a high bar, route borderline ones to the model, refuse the rest. Because the threshold is per intent class and adapts over time, the bar a creative query must clear differs from a factual one's, confidence isn't one global knob, it's a per-question judgement you can tune.

The bottom line

Reuse without a confidence number is faith. Crowkis replaces the faith with a measurement and hands you the dial, so the cache's aggression is your decision, made per hit, with the evidence attached.