Many multilingual RAG setups break in familiar ways: embedding spaces that do not line up, chunks cut mid sentence, and translation that mangles names or IDs. We tried a different path. We keep the reasoning in one base language, translate only at the edges, and guard important terms with glossaries.