Picture a visitor who writes to a shop in Thai and gets a fluent Thai answer. Around that answer sit an English "Send" button, an English form, an English "Thinking…" and an English error notice. The chatbot speaks Thai. The product around it does not.
Language support in AI products tends to stop at the model's reply. Gydr 3.0 treats language as everything the visitor reads, and the releases after it taught the assistant to handle mixed-language chats, where one message moves between languages.
Two kinds of text
Every chat window holds two kinds of text. The first is what the AI writes: the answer to the visitor's question. The second is what we write: buttons, form questions, progress messages, rating prompts, system notices, the sentence that says a person is joining.
The two need different handling. The AI's reply should follow the visitor, in whatever language they write. The sentences we author should never be improvised. A notice that someone from your team is joining, or that a message could not be delivered, has to say exactly what is true, every time. So we write those sentences ourselves and translate them ahead of time, rather than asking a model to compose them on the fly.
With Gydr 3.0, both kinds of text follow your visitor across fifty languages.
Right to left, and right from the start
For Arabic, Hebrew, Persian and Urdu, the whole chat window mirrors, so the layout reads in the same direction as the text.
The visitor's language is recognised from their very first message. That includes a refusal or a notice sent before the conversation properly starts, and a returning customer who comes back to a messaging channel after a break.
Your content does not need translating first
Most businesses write their knowledge once, in one language. In July, with Gydr 2.3, we made the assistant answer in the visitor's language even when your knowledge base is written in another. A shop with an English catalogue can serve a Chinese-speaking customer without maintaining a Chinese copy.
The hard cases
Real conversations are messier than test cases, and the releases after 3.0 went after the messy ones.
People mix languages. A visitor writing in Malay may name a product or a place in English. In Gydr 3.2.3 we made sure a foreign brand or place name does not drag the reply into another language. The reply stays in the language the visitor is actually writing.
Some languages are close relatives. Malay and Indonesian share much of their vocabulary, so in Gydr 3.2.4 we taught the assistant to tell them apart reliably. In the same release it began recognising thirteen more languages that had been answered in English until then: Italian, Dutch, Polish, Turkish, Czech, Danish, Finnish, Hungarian, Norwegian, Romanian, Swedish, Swahili and Tagalog.
Later features were built on the same rule from day one. Data Forms are asked in whatever language the visitor is writing, with no per-language setup.
Why it matters to us
Gydr was built in Malaysia, where an ordinary conversation can move between Malay, English and Chinese in a single sentence. Multilingual customers are not an edge case for our customers. They are the customer. An assistant that answers well in one language and clumsily in the rest is not good enough.
Frequently asked questions
What is a mixed-language chat?
A conversation where the visitor writes in more than one language, such as Malay with English product names. Gydr replies in the language the visitor is actually writing, so a foreign brand or place name does not switch the reply.
How much of the chat window is translated?
All of it, in fifty languages: the replies, buttons, form questions, progress messages, notices and system messages. Arabic, Hebrew, Persian and Urdu also mirror the layout right to left.
Do I have to translate my knowledge base?
No. Since Gydr 2.3 the assistant answers in the visitor's language even when your content is written in another one.
The full list of changes is in the 3.0.0 release notes.