Skip to main content
The Sandbox is your agent’s control room and test chat in one. As its subtitle says, it’s where you “adjust your agent’s settings and preview responses in real time” — set the model, temperature, instructions, and chat options, then talk to your agent exactly the way a visitor would, with a live preview beside the settings. Open Agent → Sandbox from the agent sidebar.

​
Prerequisites

​
What’s on the Sandbox

The Sandbox with the settings panel and live chat preview.
Changes you make in the Sandbox apply to your live agent once you select Save Changes. Use Undo and Redo to step through edits before saving.
The Sandbox preview is a real chat session, so test messages consume credits just like a live visitor conversation. To try lots of configurations without spending credits, use the Playground instead — its message sends are free. See Credits.

​
Let visitors speak their questions

The Allow voice input in chat toggle lets visitors record a spoken question instead of typing. As its description says, it lets users “record voice messages (up to 30 seconds) that are transcribed and sent as questions.” The recording is turned into text and sent as a normal message, so the agent answers it just like anything typed. This toggle is on by default.
  • Recording length: each voice message can be up to 30 seconds.
  • Cost: a successful voice message costs 1 credit for the transcription (on top of the normal credits the agent’s answer uses). Empty recordings — where nothing intelligible was captured — aren’t charged. See Credits.
  • Requirements: the visitor’s browser needs microphone access, and the page must be served over HTTPS. The embed snippet already requests microphone permission for the widget, so you don’t need to add anything.
Turn the toggle off for any agent where a voice option doesn’t fit. What visitors see when they use it is covered in What your visitors experience.
Sandbox setting 'Allow voice input in chat' toggled on, with its up-to-30-seconds description

​
Test your agent

1

Ask a real question

Type a question your knowledge base should answer.
2

Watch the answer stream

The agent replies in real time. While it works, you may see it thinking and using tools before the final answer appears.
3

Check the sources

Confirm the answer cites the right source from your knowledge base — that’s how you know it’s grounded in your content.
4

Adjust and retry

Tweak the model, temperature, or instructions, then ask again to compare.
Testing in the Sandbox is for your own iteration — dial the agent in here before you embed it. Because the preview runs the real chat pipeline, each test message uses credits like a live conversation; the Playground is the credit-free place to compare many setups.

​
Tips

  • Test the hard questions — edge cases, ambiguous phrasing, and things just outside your content — to see how the agent handles them.
  • If an answer is wrong, check whether the source is Completed, then sharpen your instructions or add a Q&A.
  • To compare two full configurations at once, use the Playground.

​
Troubleshooting

Make sure a relevant source shows Completed in the Knowledge Base. If it does, add a closer-matching Q&A pair.
Refine your instructions and consider lowering temperature.

Model & temperature

Instructions

Playground

Embed your agent