How many Australians use another language at home?
The 2021 Census counted 5,663,709 people who used a language other than English at home, and 852,706 of them reported that they do not speak English well or at all. Mandarin was the most common language other than English used at home, followed by Arabic, and Punjabi showed the largest increase since 2016. For aged care providers, health services, councils and community organisations, these figures describe real enquiries. They also show why the supported language list matters: a model that handles French well may not help a family that speaks Punjabi.
How do you check which languages a model supports?
Read the model card instead of the marketing page. Model cards usually state which languages the developer evaluated for quality and safety. Meta's Llama 3.1 card lists eight supported languages: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. It notes the model was trained on a broader set of languages, but it lists use in other languages as out of scope and strongly discourages it without fine-tuning and system controls. A model can produce fluent-looking text in a language it was not evaluated for, so a fluent demo proves little. Record the supported list for every candidate model beside your required languages.
Why does the search step need a multilingual embedding model?
In a retrieval system, the question and the documents are converted to embeddings and compared. If the embedding model only handles English, a question typed in Vietnamese may not match the English policy that answers it, and the chat model never sees the right source. Multilingual embedding models place different languages in a shared space. Cohere, for example, says its embed-multilingual-v3.0 model supports over 100 languages, and its list includes Chinese, Arabic, Punjabi and Vietnamese, while some of its other embed models are English only. Check the embedding model's language list as carefully as the chat model's.
Should you translate the documents or the answers?
There are two common designs. In the first, documents stay in English, the system searches across languages and the model answers in the user's language. In the second, you maintain translated versions of key documents, reviewed by qualified translators, and the system answers from the version in the user's language. The first needs less upkeep because there are no translations to maintain. The second gives you control over wording where precise meaning matters, such as fees, consent, care plans or safety instructions. You can combine them, with approved translations for a short list of critical documents and cross-language search for everything else.
When is a machine answer not good enough?
Machine translation and LLM answers can miss nuance, register and cultural context, and errors are harder to spot when staff do not read the language. For clinical, legal, financial hardship and safety-critical conversations, use qualified interpreters and translators. The Translating and Interpreting Service (TIS National) is provided by the Department of Home Affairs for people who do not speak English and for agencies and businesses that need to communicate with them. NAATI is Australia's national standards and certifying authority for translators and interpreters. Use the LLM for routine information and make the route to an interpreter obvious.
How do you test answers in another language?
Recruit native speakers, ideally staff or community partners who know your services, and give them real questions in their language. Ask them to judge separately whether the answer is correct against the English source, whether the language is natural and respectful, and whether the citation helps. Include questions that should be refused or handed to a person. Hypothetical example: three target languages, 30 questions each, reviewed by two speakers per language; the counts are illustrative only. Record failures by language, because quality can differ between languages within the same model.
Which multilingual tasks are sensible to start with?
Start where a wrong word is easy to notice and cheap to fix. Routine service information is a reasonable first step: opening hours, how to book, what to bring to an appointment, how to update contact details. A second option is staff-facing help, where a bilingual employee asks the assistant to summarise an English procedure in another language and checks the result before sharing it. Leave medication instructions, legal rights, consent and complaints with qualified people until you have test results for that language and content type.
- Routine public information drawn from pages you already publish in English.
- Draft summaries for bilingual staff to check before use.
- Finding the right English document from a question asked in another language.