-
Notifications
You must be signed in to change notification settings - Fork 0
Gen3 Chat Starting Conversations
- Open GT Chat from the sidebar or launch from an agent card.
- Choose the agent that matches your workflow.
- Optionally attach datasets before the first message.
- Send an opening prompt to create the conversation thread.
Starting conversations correctly sets agent and dataset scope up front, which keeps retrieval and exports predictable later.
Use this page when you want the most direct path into live work in GT Chat.
Go straight to the GT Chat route when you already know which agent and datasets you want to use.
Use Agents when the first decision is which agent should own the work. This is the best path when you maintain favorites or when you have just updated an agent and want to test it immediately.
You can resume prior work from the chat history rail or from the broader Conversations page when you need search and filters first.
Confirm all three of these items:
- The correct agent is selected.
- The conversation has the required dataset context.
- Any newly uploaded files have finished processing.
If any of those checks are wrong, fix them before you send the prompt. It is easier to start with the right context than to debug a weak answer afterward.
| Situation | Best path |
|---|---|
| You already know the agent and want to start now | Open GT Chat directly |
| You need to compare or choose an agent first | Start from Agents |
| You need to continue earlier work | Reopen from chat history or Conversations |
- Start a new conversation when the work changes meaningfully.
- Continue an existing conversation when prior context still matters.
- Rename or otherwise preserve important work through the history workflow on Conversations instead of relying on memory later.
- Attach datasets before the first prompt when source-grounded answers matter.
If the first answer is not grounded the way you expected:
- confirm the agent was the intended one
- confirm the right datasets were attached
- wait for upload processing to finish if you just added files
- check whether the source documents actually belong in Datasets before retrying