Translation¶

Backend.AI GO translates in ordinary chat. Ask for a translation in a message, and attach a document or an image when you want a file translated. The translation runs on the model you chat with, so with a local model your text and files never leave your machine.

Translate text¶

Ask for a translation in a message, for example "Translate this into English: ...". The chat model opens the translation view: the conversation splits into the source on the left and its translation on the right. The text can also be one already in the conversation, such as the model's last answer ("Translate that into English"): the model copies it into the view as the source.

• Language: when you name a language, the view translates into it right away. When you do not, the right pane stays empty until you choose one from the language list at the top of the view.
• Editing: the source is editable. When you stop typing, only the sentences you changed are translated again, and each new translation is highlighted for a moment. The rest of the translation stays as it was.
• Sentences in context: each sentence is translated with the sentences around it, so pronouns and terms follow the rest of the text. A sentence's translation appears as the model writes it.
• Translation model: the view starts with the chat model. The Translation model picker at the top right of the view changes the model for this translation only, while the chat model in the composer below stays as it is and answers your messages. A downloaded model that is not running starts when you pick it.
• Glossaries: a glossary whose terms appear in the source applies to the translation, as it does in chat.
• Copy and download: the copy button under the translation copies it, and the download button beside it saves it as TXT, Markdown (MD), or Word (DOCX). The file is named after the source file, or "translation" for typed text, with the language code added (notes.en.md, translation.ja.docx). A Markdown download keeps a Markdown file's structure.

A request alone, such as "Let's translate", opens an empty view to type into. A model that cannot call tools, such as TranslateGemma or a small local model, cannot open the view itself, so click + in the composer and choose Translate. That entry works with every model, and the text already in the composer becomes the source.

Keep chatting while the view is open¶

The composer stays under the view, and a message sent from it goes to the chat model with the translation. Ask for a change such as another language ("Change it to Japanese") or a style ("Make it more formal"), and the view stays open and updates. Any other message gets a normal reply, and the view folds into a card in the conversation. The close button at the top right folds it the same way.

A style applies to the whole translation, so every sentence is translated again in it. It shows at the top of the view as a chip, for example Style: more formal ×. Click the chip to translate without the style. A dedicated translation model such as TranslateGemma translates each sentence alone and does not follow a style.

The card shows the language pair and the start of the source. Click it to open the view as you left it, with the source, the translation, the language, the style, and the model. The translation is saved with the conversation, so the card still opens it after you restart the app.

Translate a document or an image¶

1. Click + in the composer and choose Attach file, or drag the file onto the composer.

   • Documents: TXT, Markdown (MD), Word (DOCX), PDF, and EPUB, up to 10 MB. A document attaches whichever model you use, because it goes to the translator rather than to the model.
   • Images: PNG, JPG, GIF, and WebP. An image attaches when the model can read images.

2. Ask for the translation, for example "Translate this file into English".

3. The model calls a translation tool for the file:

   • TXT or Markdown: the file opens in the translation view with its text on the left, which you can edit like typed text. A Markdown file keeps its headings, lists, and quotes, and its code blocks stay untranslated.
   • Image: the image opens in the translation view, with the text read from it under the picture. Click the picture to see it larger. Correct the text if it was misread, and only the sentences you changed are translated again.

   • Word, PDF, or EPUB: the translator translates the whole file and the translation view follows it. The files of your message are listed on the left with their type, size, and segment count. On the right, the selected file shows its progress while it is translated, then the result. Files are translated one at a time, in the order the model asked for them.

     Cancel translation under a file's progress stops that file once you confirm. The segments translated so far are discarded and no translated file is made. A file still waiting for its turn can be cancelled the same way.

     A translation keeps running when the page reloads, for example in a browser. When you open the conversation again, the view shows the file's progress, and its result card appears when it finishes. A translation does not survive quitting the app or restarting the server: the conversation then says it stopped before it finished, and you can ask again.

   An image with no text to read gets a card with the result instead.

4. For a Word, PDF, or EPUB file, use the result in the view:

   • PDF or EPUB: the translation is shown as a document. The copy button copies it, and the download button saves it as TXT, MD, or DOCX.
   • Word: there is no preview; the translated file keeps the original's formatting.
   • Every format: the translated file is named after the original with the target language added (report_en.docx). In the desktop app, Show in folder shows it in Finder or Explorer. In a browser, Download saves it.

   The view folds into a card for each file when you send another message and the model answers normally, or when you close it. View translation on a card opens the view again with the result, also after you restart the app.

PDF and EPUB translations are saved as Markdown (.md). EPUB extraction follows the chapters' reading order and keeps their headings, while the original layout and images are not rebuilt. A Word document keeps its structure.

A file is translated by a tool the model calls, so the model has to support tool calling and tool calling has to be on (see Tool Calling). When either is missing, a notice above the composer says the attached document will not be translated. Text you type still translates in the reply.

Glossaries¶

A glossary fixes how specific terms are translated, such as product names, technical terms, and domain vocabulary. Glossaries live in the Glossary tab of Data, between Documents and Wiki.

Glossaries apply on their own¶

When a term from a glossary appears in your message, or in the text of what you attached (a TXT or Markdown file, or a document attached from Data), that glossary applies to the message. A chip above the composer names it, for example Glossary: Product terms ×. The matching terms go to the model with your message, and the translation tools use the same glossary for the file.

Click the chip to turn that glossary off for the message. Matching only looks for the glossary's source terms in the text; it does not try to guess whether you asked for a translation. A term written in Latin letters matches as a whole word, so "cat" does not match "category", while a Korean term also matches with a particle attached ("제품을" matches "제품").

Pick a glossary by hand¶

Some terms cannot be found before the message is sent: the words in an image, or in a Word, PDF, or EPUB file. For those, click + in the composer, choose Use a glossary, select one or more glossaries, and click Apply. A glossary picked this way applies to the next message in full.

Manage glossaries¶

Open Data and its Glossary tab, or run Go to Glossary from the command palette. Each glossary card shows its name and description, its source and target languages, and its number of terms, with these actions:

Creating a glossary¶

1. Click Create Glossary.

2. Fill in the glossary information:

   • Name (required): A descriptive name for the glossary.
   • Description: Optional notes about the glossary's purpose.
   • Source Language (required): The language of the original terms.
   • Target Languages (required): One or more languages to translate into.

3. Click Create Glossary in the editor to save it.

Editing a glossary¶

Click Edit Glossary to open the editor, where you can:

• Add new term entries with the Add Entry button.
• Define source terms and their translations for each target language.
• Remove entries you no longer need.

Compatible models¶

Any chat model translates text in the translation view: a model that calls tools opens it from your request, and the + menu opens it for any other model, which makes it the place to use a dedicated translation model such as TranslateGemma. To translate an attached file, choose a model that supports tool calling; the translation itself runs on that same model, and an image needs a model that can read images. Dedicated translation models such as TranslateGemma do not call tools, so attach files while chatting with a general model that does.

Organization policy¶

An organization that hides the page id /translation hides every part of translation listed here: the translation view and its Translate entry, the translation tools, the glossary chip and the Use a glossary entry, and the Glossary tab in Data. Text can still be translated in a reply, since that is ordinary chat. See Policy Schema Reference.

Tips¶

• Create a glossary for each project or domain, so terms match the glossary that owns them.
• Translation in chat works with local models, so confidential documents can stay on your machine.