Chatbox AI Assistant

Chatbox is the conversational side of Transbox — for the questions a translation engine cannot answer: how a word should read in this context, whether a rendering sounds natural, what differs between two options.

It shares the same model layer with the in-editor assistant: the AI assistant tab in the editor is the same chat engine, with a session scoped to each segment.

Where Models Come From

• Platform models: hosted by the platform, no setup needed, billed in credits by usage
• Custom models: any OpenAI-compatible service (model ID, Base URL, API key, optional custom headers). Requests go directly from your machine to the provider

Interface

Session list (left)

• New session at the top
• Search filters sessions by name
• Each session shows its name and the active marker
• Sessions are ordered by creation; switching sessions never interrupts saving of existing records

Session controls (top)

• Name: click to rename; the refresh button regenerates it from the latest question/answer pair
• Model: switch the model for this session; the change applies immediately and is stored on the session
• Continuous conversation: when on, each message carries the session's recent history (up to 10 messages); when off, every message is standalone
• System prompt: set the role and constraints for this session (e.g. "You are a software localization translator; keep the target concise")
• Delete session: removes the session and all its messages

Messages

• Enter sends, Shift + Enter inserts a line break
• Stop interrupts generation and cancels the upstream request
• Messages appear chronologically; scroll up for history and use the "back to latest" affordance when new content arrives

Use Cases

Translation questions

• Idioms, slang and culture-bound terms
• Choosing between renderings of a term in different contexts
• Splitting long sentences and reordering information

Polishing

Paste a machine translation, ask what reads unnaturally and for a rewrite, then apply the change back in the editor.

In-editor assistance

In the project editor, the AI assistant tab:

• Sessions are isolated per segment + target language, so switching segments switches context
• The project-level AI assistant prompt acts as the system prompt for every segment in that project
• Ask for candidate translations, then decide together with memory and termbase evidence

Tips

1. Give context: state the text type (UI string, legal clause, marketing copy) and the audience — the quality gap is large
2. Constrain the output: e.g. "give three options with the context each fits, nothing else"
3. Control continuity: turn continuous conversation on for follow-ups, off for one-off questions
4. Fix style in the system prompt: for long-running projects, put the style rules there once

Data and Cost

• Local storage: sessions and messages live in the local database and travel with Data backup
• Cost: platform models consume credits; custom models are billed by your provider directly, Transbox never touches the payment
• Privacy: with platform models, content goes to the platform service; with custom models it goes straight to the provider you configured

FAQ

Q: How is Chatbox different from text translation?

A: Text translation returns a single result. Chatbox lets you ask follow-ups, get explanations, compare renderings — useful when you need to reason about a choice.

Q: Are the built-in models free and unlimited?

A: No. Platform models consume credits depending on the model. MT engines are currently free for a limited time; the two have different billing.

Q: How long are conversations kept?

A: Indefinitely, until you delete the session. Nothing expires automatically.

Q: Can I export a conversation?

A: Not currently. Copy the message text if you need to keep it.

Q: Why do answers change so much after switching models?

A: Models differ in capability and default tone. Use a system prompt to hold the style steady across models.

Q: Can I use AI output as the delivered translation?

A: Not directly. Models can produce plausible but wrong terminology or meaning; review before delivery, especially for legal, medical or financial content.

Q: Do the editor's AI assistant and Chatbox share history?

A: No. The editor's assistant creates per-segment sessions independent of the Chatbox session list, but both use the same models and system-prompt mechanism.