Skip to content

CONCEPT Cited by 1 source

Semantic-cache precision-recall threshold

A semantic-cache precision-recall threshold is the minimum similarity score at which a vector-retrieved cached artifact is trusted as reusable. It controls a different risk than ordinary cache eviction: it determines whether two differently worded requests have sufficiently similar intent and structure to share an executable template. (Source: sources/2026-08-13-aws-reducing-text2sql-latency-with-parameterized-query-templates)

Operating trade-off

Threshold posture Retrieval result Consequence
Too strict Rejects valid paraphrases More authoritative LLM generation, lower cache hit rate, higher cost and latency
Too loose Admits a related but structurally wrong template Incorrect query or incomplete answer can pass the fast path
Calibrated Reuses only safely similar templates Captures cache economics while preserving fallback for ambiguity

The correct value depends on domain narrowness, template coverage, and the cost of a false match. AWS recommends recording selected templates and scores in production, examining both rejected good matches and accepted bad matches. (Source: sources/2026-08-13-aws-reducing-text2sql-latency-with-parameterized-query-templates)

Retrieve broadly, decide narrowly

When one similarity score is not precise enough, first retrieve a wider candidate set at a looser threshold and rerank with a small LLM or specialized reranker. This preserves recall while moving the accept/reject decision to a stronger selector; it does not eliminate the need for result sufficiency checks and authoritative fallback. (Source: sources/2026-08-13-aws-reducing-text2sql-latency-with-parameterized-query-templates)

Seen in

Last updated · 622 distilled / 1,953 read