Multilingual AI Knowledge Base: One Truth in Every Language Your Customers Speak

A multilingual AI knowledge base is not one knowledge base with a translate button on top. It is a set of language versions of the same truths, organised so the agent answers a Spanish question from the Spanish documents, a Portuguese question from the Portuguese documents, and never blends the two into a half-translated reply. The failure mode is familiar: the English price list is current, the Spanish one is two seasons old, and Spanish-speaking customers get last year's prices with this year's confidence. Organisation prevents that.

One intent, one document per language

Mirror the structure across languages: if English has a delivery document, a returns document, and a price list, Spanish and Portuguese each get the same three documents, titled in their own language and covering the same facts. Mirroring makes gaps visible at a glance — a language missing the returns document is a language whose customers cannot get a returns answer — and it keeps retrieval inside one language, where answers stay fluent. An illustrative example: a tour operator serving English, Spanish, and Portuguese travellers keeps three parallel document sets for tours, pickup points, and cancellation terms; a customer asking in Portuguese retrieves only Portuguese passages, so the answer never arrives half in English.

Translate intent, not just words

Machine translation of a document is a starting point, not a finished knowledge base. Idioms, formality levels, and local terms — the name customers use for a neighbourhood, the word for a booking deposit — differ by market, and a literal translation that nobody would say retrieves poorly for the questions people actually ask. Have a speaker of each language review the translated documents for the terms customers use, and rewrite the question phrasings in each language from real chats rather than translating the English questions. The knowledge is the same; the language it wears must be local.

Mixed-language questions and the fallback rule

• Answer in the language the customer used, even if the question mixes languages — a customer writing mostly Spanish with an English product name still gets a Spanish answer.
• Set one fallback language for questions in a language with no documents yet, and say so in the reply: “I don't have that in your language yet — here it is in English, and our team can confirm in your language.” A declared fallback beats a guessed translation.
• Never let the agent answer from a different language's document and translate on the fly without saying so — on-the-fly translation of prices, dates, and conditions is where silent errors enter.
• Log unanswered questions by language. A cluster of unanswered questions in one language is a document to write in that language, not a reason to route everyone to English.

Keeping every language in sync

Translations go stale the moment the source changes, so treat every edit as a multi-language edit: when the English price list changes, the Spanish and Portuguese versions change in the same session, by the same person or team, before the agent answers another question. Keep a simple sync sheet — one row per document, one column per language, with the date each version was last updated — and review it monthly. A language version more than one change behind the source is a wrong-answers risk, and the sheet is what makes that visible before customers find it. On Neksio, Facebook and Instagram channels sit on the Growth plan at $11.99 a month for 15,000 replies, while the Free plan's 100 replies a month cover WhatsApp only — so a multilingual rollout across channels should be sized on the plan that actually carries those channels.

Owner: “Our Spanish customers kept getting old prices.” The English list had been updated twice; the Spanish translation had not been touched since launch. One sync sheet and one rule — no source edit ships without its translations — ended the problem.