A visitor writes in Spanish. The answer comes back in English, or half in each, or the buttons are in one language and the text in another. Somebody forwards it with the subject "this is broken".
It usually is not broken. Three separate things decide the language of a reply, and the symptom tells you which one is misconfigured. Retrieval, the part everybody suspects first, is almost never the problem.
Find your symptom
The three dials, and what each one controls
The fourth thing people reach for, a separate knowledge base per language, is not on that list on purpose. It is almost never the fix, and the next section is why.
Retrieval is not your problem
This is the part that surprises teams: your Spanish question finds your English documents. A multilingual embedding model puts meaning in the same place whatever language carries it, so "¿puedo devolverlo?" lands beside the page that says "Returns and exchanges".
Which means you do not need to translate your help centre to answer Spanish questions. Write the policy once, and the retrieval half works across languages on day one.
welcomeai.devWhat an embedding is, and why two sentences with no shared words land togetherThe instruction that fixes the most common case
If answers come back in the wrong language, the instructions are usually doing it, and the fix is one explicit line rather than a setting.
Write it plainly: always reply in the language the visitor wrote in, even when the source material is in another language, and never apologise for the source being in English. Then add the exceptions that matter in your business: the product names, plan names and legal terms that stay as they are in every language.
Ask the same question in each, and read the whole reply rather than the first line. The mixing problem hides at the end of an answer, in the sentence that was lifted from the source.
When a separate knowledge base per language does earn its place
Three cases, and none of them is "we have Spanish customers".
Legally binding wording, where the translated version is the one that governs in that market. Content that genuinely differs by country, where prices, shipping or warranty terms are not the same text in another language. And markets where you want a different tone or a different set of examples, not a translation.
Everywhere else, one knowledge base and one clear instruction will do better, because two bases mean two things to keep current and one of them will quietly fall behind.
Questions people ask
Pick the symptom in the embed, make the one change it points at, and ask the same question again. Nearly every report of "the chatbot speaks the wrong language" is one line of instructions or one attribute on a tag.
welcomeai.devAI chatbot for multilingual websites
