Support Multiple Languages with Your AI Assistant
The AI Assistant answers in the reader's language. The OpenAI models it runs on read and write dozens of languages, so readers ask in their own language and get answers in it — even when your documentation is written in another. This topic explains what to expect from each language, how the Assistant handles questions and documentation in different languages, how to structure Assistants for a multilingual help site, and how language affects usage.
What to Expect from Each Language
Answer quality in a language tracks how much published text in that language the underlying model was trained on:
English is the strongest language.
Widely published languages — Spanish, French, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Russian, Arabic, and their peers — perform close to English, with fluent and accurate answers.
Languages with a small published footprint taper: answers stay understandable, but phrasing stiffens and comprehension of subtle questions weakens.
The floor is high and rising. In OpenAI's 26-language benchmark, the model scored better in 24 of them — including Latvian, Welsh, and Swahili — than its predecessor scored in English, and each generation since has narrowed the remaining gaps. Two practical consequences:
Model choice matters most outside the major languages. The more capable models hold up noticeably better there — weigh that when you choose a model, and compare candidates in the Preview tab with questions in the languages your audience uses.
Language coverage is a property of the model, not your setup. Switching an Assistant to a newer model upgrades every language at once.
Ask in One Language, Answer from Another
The Assistant detects the language of each question and replies in it. There is no setting to configure: a reader who asks in French gets French back, whatever language your Instructions and documentation use. The model reasons internally in English, so any streamed reasoning summary shown by the thinking indicator may appear in English; the answer arrives in the reader's language.
The documentation's language does not have to match the question's. The model matches meaning across languages: it answers a German question from English help pages and translates as it goes. Two behaviors to design around:
Lookups are most reliable when the languages match. A question phrased in the same language as the content finds the right passage more consistently than a cross-language lookup. When most questions will arrive in a given language, provide help content in that language.
Names stay as your documentation spells them. Setting names, menu labels, and values are quoted the way your pages write them: a Spanish answer drawn from English documentation names the Generate Assistant setting in English, matching what the reader sees in the product.
Keep exact product terms, setting names, and UI labels identical in every language you publish. Consistent names are what tie a question in one language to content in another.
To pin the behavior, state it in the Instructions (see Write Instructions That Hold Up): "Always answer in the language of the reader's question," or, for a single-language site, "Always answer in French, whatever language the question uses."
Structure Assistants for a Multilingual Site
How you organize Assistants depends on what "multilingual" means for your site:
The same content, translated
When each language is a translation of the same help, create one Assistant per language. Each Assistant gets its own Knowledge Base (the knowledge-*.zip from that language's Reverb 2.0 target) and Instructions written for that audience. Every lookup stays within one language — the most reliable configuration — and every answer links to that language's pages. This matches the localization workflow of one target per language (see Specifying the Language Used by Targets) and the rule of one Assistant per target (see Create an Assistant on the Platform).
Different content per language
When the languages carry different information — a general manual in English plus region-specific topics in the regional language — use a single Assistant with the mixed-language Knowledge Base. Terminology continuity binds the languages: keep the shared product names, setting names, and keywords identical in every language so a lookup can travel between them.
Upload each piece of content in one language per Assistant. The same content in several languages splits lookups between the copies and makes answers less consistent; translated duplicates belong in separate per-language Assistants.
How Language Affects Usage
Usage is measured in tokens — the word-fragments the model reads and writes. Token counts vary by language. English packs the most meaning into the fewest tokens; major European languages run modestly higher, both from tokenization and from natural verbosity (the same answer takes more words in German than in English); other scripts — Japanese, Arabic, Hindi, Thai, and similar — run higher still. Plan for a non-English audience to consume usage somewhat faster than an English one at the same question volume. The same applies to what the model reads: content in a verbose language consumes context faster during lookups.
Test Every Language You Support
Extend the question list from Test with a Question List to each language you expect questions in:
Write the questions natively. Have a fluent speaker phrase them the way a real user would. Machine-translated questions mirror the documentation's own wording and miss how real users ask.
Cover the cross-language case you have. When readers ask in one language and your content is in another, most of the list should do the same.
Check the links. Each answer's source link should land on a page in the reader's language — on a per-language Assistant, always its own language's pages.
Include trick questions per language. An invented setting name should get the same "I'm not sure" in every language.
Judge fluency with a native reader. An answer can be factually right and still read poorly; only a native reader catches it.