Efficiency Is Not Language-Neutral: Per-Language Auditing and Calibration of Multilingual Dense Retrieval

Production dense retrieval trades accuracy for efficiency: vectors are compressed, search is approximate, and thresholds filter results, each tuned by mean quality. In a multilingual index, one configuration serves every language, so losses on a few can hide behind a healthy mean. This practice-oriented empirical study audits that efficiency layer per language, separating hidden losses from sampling noise and generic fragility, across eight embedding models, three test beds, indexes of up to 4,978,232 passages, and five vector databases. The mean hides the worst-served language beyond noise in 339 of 474 compression configurations; with BGE-M3 and 16-byte product codes, mean retention is 0.767, but romanized Bengali keeps 0.228. Vector databases at documented settings hide it in 26 of 27 settings with a mean retention of at least 0.95, and the loss reaches downstream answers. Trained product quantizers concentrate error beyond matched noise, whereas dequantized int8 is nearly lossless for every language. The study offers a reusable audit protocol, a per-language map of safe configurations, and per-language calibration with finite-sample guarantees that keeps every language within target in all 464 settings; marginal and per-query calibration fail in 325 and 300. Efficiency in multilingual retrieval should be measured, and calibrated, per language.