Token overlap

vs word-overlap (Jaccard)

Treat high token overlap as “the same question.” Cheap, local, and noisy, especially on freshness.

Asked

latest node release?

Matched on shared words

node release notes

Shared words, different question

Wrong hit

Where MorrowCache wins

Intent adjudication plus freshness refuse. Offline Jaccard@0.35 on n=103 showed false-positive rate 0.49, that number is the overlap cache, not Morrow. Overlap alone is not a safe primary gate here.

Where they win

Overlap is local, fast, and needs no model judge. For toy demos or ultra-constrained vocabularies it can look fine. It does not know “what’s the latest,” and it will happily hit the wrong FAQ.

Verdict

We measured overlap and rejected it as the primary gate.

Get started Full agent guide

Other comparisons

Always calling the model
Exact-only cache