Playground

Overview

The Overview sub-tab is a control panel plus a live chat, side by side. On the left: the model picker (switch models instantly, no save needed to try one), a source count with a shortcut into Data sources, a quick-edit box for instructions that mirrors the full editor, and a visibility toggle — turn it off and the chat bubble channel stops accepting messages entirely, on every site it's embedded on.

A "How this works" card explains the retrieval pipeline in the same order it actually runs:

1. Your question is embedded and matched against this agent's chunks only, by cosine similarity.
2. The top 5 most relevant chunks are inserted into the model's context.
3. The model answers only from those sources, or says it doesn't know.
4. Every citation links back to the exact source chunk and its similarity score.

Messages you send here run through the exact same pipeline a real visitor's message would — same model, same retrieval, same instructions, same guardrails — so what you see in the Playground is what ships. The one difference: Playground conversations aren't rate-limited the way the public widget is.

Display

Everything a visitor sees. All of it saves to the agent and applies immediately to every real channel that embeds the widget (chat bubble, help page, Center Stage) — there's no separate "publish" step.

• Theme — light or dark, plus a custom accent color used for the header and buttons.
• Corner radius — none, small, medium, or fully rounded, applied to the whole widget shell.
• Floating bubble — left or right corner placement, an icon style, a custom button color, and whether the bubble shows a text label alongside the icon.
• Visibility — the same toggle as Overview, surfaced here too since it's a display concern.
• Voice to text (dictation) — adds a mic button to the widget's input box that uses the visitor's own browser speech recognition. Support depends on their browser, since nothing is sent to a third-party speech API for this.
• Delete conversations — lets a visitor clear their own current conversation from inside the widget.

A live preview panel re-renders as you change any of these, so you're styling against the real widget, not a mockup.

Voice

Configuration for the agent's speech behavior — separate from Settings, which covers text-channel behavior like rate limits. See the Channels page for how this connects to the real Phone channel, which answers inbound calls over a Twilio voice number using this same agent.

Actions

A test surface for the tools you've built on the Actions tab. Trigger one manually with sample arguments and see the exact request that would go out and the response that comes back, without needing a real conversation (or a real visitor) to exercise it.