Glossary

Very close (Jaccard similarity)

Automatic marker for a prompt deemed a near-duplicate of another already in the batch, based on Jaccard similarity computed on tokens (normalised words). Typical threshold: > 0.85.

Concrete example

"What are the best portable ACs for a studio?" and "Best portable AC for a small apartment?" share most of their tokens: the second is flagged "Very close" to the first.

Why it matters

Without this guardrail, a narrow pool would produce artificially inflated batches. The marker lets you clean the batch without losing credits and hints that the pool may need broadening.

Related terms