Closing the Calibration Gap in Semantic Caching
For practitioners deploying semantic caching to reduce LLM inference costs, this work reveals a fundamental metric mismatch that causes suboptimal model choices, shifting focus from ranking to calibration.
The paper identifies that PR-AUC, the standard metric for semantic caching, misaligns with deployment performance, leading to poor model selection. It proposes two new metrics—P-CHR AUC and CRR—and shows that the calibration gap is driven by training objective, not data scale, and post-hoc calibration only partially fixes it.
Semantic caching cuts LLM inference costs by serving a cached response to semantically similar queries. Standard practice evaluates these systems using PR-AUC, a metric that only measures how well scores rank and ignores whether they are usable at a fixed threshold. We show this mismatch leads to systematically poor deployment choices, as models with the highest PR-AUC are often the worst in operation. We introduce Precision-Cache Hit Ratio (P-CHR) AUC, a cache-aware metric that measures precision across cache utilization levels, and Calibration Retention Rate (CRR), which captures how much offline ranking quality survives at deployment. We decompose the operational gap between offline and deployed quality into a recoverable calibration component and an irreducible structural component fixed by the dataset's positive rate. Our experiments show that the calibration gap is governed by the training objective rather than data scale, and post-hoc calibration only partially closes it. Ultimately, model selection for semantic caching is a calibration problem, not a ranking one, and measuring it is the first step to closing the gap.