Gen3 Chat - GT-Edge-AI/GT-AI-OS GitHub Wiki
GT Chat
Start Here
- Open GT Chat and select an agent.
- Tune layout and density in Chat Window Preferences (Chat Preferences); set light/dark appearance from the account menu (Accessibility).
- Attach datasets or upload files if the task needs source material.
- Wait for document processing to finish before the next prompt when uploads are pending.
- Send your prompt and review the streamed response.
- Use copy, download, Create Image, voice input, or report actions on assistant messages when needed.
Why this matters
GT Chat is the live execution surface—agent, conversation, datasets, uploads, and streaming answers in one workspace.
Details
GT Chat is the live execution surface for the current Gen 3 tenant app. It combines the selected agent, the current conversation transcript, any conversation-attached datasets, staged uploads, and the assistant response stream in one workspace.
What the active GT3 chat page does
- starts new conversations
- reopens recent conversations from the chat sidebar
- launches chat from an agent card or favorite agent
- attaches existing datasets to the current conversation
- uploads files into a chat-created or user-selected dataset
- shows response actions such as copy, download, Create Image (when image generation is enabled), Read aloud, and report
- supports Speech to Text when STT models are enabled; Hands-Free Mode additionally requires platform TTS and a resolvable agent TTS model
- allows read-only accounts to upload conversation-scoped files when a tenant owner enables Chat Upload Access on the user record
- keeps assistant message actions sticky at the top of the transcript viewport while you scroll long answers
- renders generated images below assistant text in the message body (artifacts follow the markdown answer)
- shows a Copy button on fenced code blocks in assistant replies (top-right of each code block)
Conversation history still exists outside chat, but the dedicated history route is Conversations, not a separate documents or workflow page.
Dataset retrieval vs web search in chat
Favorited agents may call search_datasets when conversation or agent datasets provide scope—that tool retrieves from your tenant corpora, not the public internet. web_search is a separate managed tool for live public-web research and requires deployment defaults plus per-agent web search settings (see Building Agents). Do not expect current-events answers from dataset retrieval alone.
Standard chat workflow
- Choose an agent.
- Confirm whether that agent already brings its own default datasets.
- Add any extra datasets or upload files for this conversation.
- Send your prompt.
- Wait for uploads to finish processing if the chat shows pending document ingestion.
- Review the assistant response and use copy, download, or report actions when needed.
Understanding the main chat controls
Agent selector
Every conversation runs through a selected agent. You can arrive in chat from the direct /chat route, or by opening chat from Agents. If you need a different behavior profile, switch agents before sending the next prompt.
Datasets panel
The Add Datasets to Conversation panel lets you attach existing datasets or upload new files into a dataset. Use it when the current conversation needs retrieval from specific source material.
Conversation history rail
The chat sidebar keeps recent conversations close to the live workspace. Use it for quick switching. Move to Conversations when you need broader search, filtering, or exports across many threads.
Assistant message actions
Completed assistant messages expose a sticky action row (configurable in Chat Preferences):
- Copy — place the rendered answer on your clipboard
- Copy (code blocks) — copy fenced code from assistant replies without selecting text manually
- Read aloud — TTS playback when deployment STT/TTS gates allow
- Download — export the single response or full conversation (see Reviewing and Exporting Conversations)
- Create Image — generate an image artifact from the assistant answer when image generation is enabled
- Report Chat Issue — capture the prompt/response pair for support investigation
Voice mode
When STT is enabled, the composer microphone supports Speech to Text (single utterance). Hands-Free Mode (continuous listen → transcribe → send → optional read-aloud loop) appears only when platform TTS is also enabled and the agent has a resolvable TTS model—otherwise the popup title is Speech to Text only and a second microphone click closes the bubble instead of entering hands-free. Status text in the voice popup shows listening, transcribing, and active hands-free states when applicable. Voice features require Platform Multimodal Settings and agent speech configuration—see Accessibility.
Read-only chat uploads
Read-only tenant users normally cannot mutate catalogs, but owners may allow chat dataset uploads per user. When enabled, read-only users can attach conversation-scoped files through the paperclip flow without gaining dataset edit rights elsewhere. Owners manage this on Users (per-user edit shelf or Bulk Chat Upload Access).
Working safely with uploads
Uploaded files do not become instantly searchable. Gen 3 blocks the next message until the staged files finish processing into the target dataset. If you upload files and the composer remains disabled, wait for the processing indicator to clear before sending the next prompt.
When to leave GT Chat for another page
| Need | Best page |
|---|---|
| Build or edit the agent itself | Agents |
| Bulk-manage datasets and documents | Datasets |
| Search or export older threads | Conversations |
| Review usage or access analytics | Observability |
Best practices
- Keep one conversation focused on one line of work when you expect to export or review it later.
- Attach only the datasets you actually need so retrieval scope stays predictable.
- Use agent defaults for repeatable work and conversation-attached datasets for one-off context.
- Report bad outputs from the message that demonstrates the issue instead of describing the problem from memory later.