Language and locale mismatches in retrieval

Multilingual embedding models place equivalent text in different languages close together: Multilingual E5 matches translation pairs across languages (bitext mining, BUCC 2018 and 112-language Tatoeba) and retrieves across 16 languages on MIRACL, so a query in one language can retrieve a passage in another. In multilingual RAG experiments (ACL Findings 2025) retrievers preferred documents in the query language and in high-resource languages, and generators preferred the query language. That an answer engine answers from the source language when no localised passage exists, and prefers a comparable localised one when it does, is inference from these results.