Playground
The Playground runs a prompt against a model and shows the output. You can open several windows side by side to compare prompts, models or settings, then save the best window to Prompts.
The Playground is not in the sidebar today. Open it with the Playground button on a prompt or on a generation in a trace, or go to /project/<projectId>/playground. Both buttons let you start a Fresh playground or Add to existing windows.
Where the models come from
| Source | When it is used |
|---|---|
| LLM Connections | The default. The model list shows each connection's default models (if enabled) and its custom models. Calls use your provider key and are billed by your provider. |
| Budget AI model groups | When Budget AI is available to you and the project has model groups, the Playground lists them under the provider Gateway. Calls go through Budget AI, so keys, budgets and routing apply. |
Running the Playground through LLM connections needs a role that can read LLM connections. Running it through Budget AI needs Budget AI keys write access. See Members & Roles.
What a window contains
- Messages. System, user and assistant messages with
{{variable}}placeholders. Chat prompts can also hold message placeholders for a list of messages, such as chat history. - Model and parameters. Provider, model, temperature, max tokens and the other parameters the model supports.
- Variables. Values for each
{{variable}}, filled in before the call. - Structured Output. A JSON schema that the response must follow. See Schemas & Tools.
- Tools. Tool definitions the model may call. The Playground shows the tool calls the model makes. It does not run your functions.
- Output. The response, which you can copy, view as JSON, or add to the messages as an assistant message to continue the conversation.
Compare side by side
Use Copy window to duplicate a window, change one thing, and run both. Run All (Ctrl + Enter) runs every window at once. You can have up to 10 windows. Window layout and contents are kept in your browser, so they survive a page refresh. Reset playground clears all windows.
A typical loop
Open
Open a prompt version in the Playground, or start from an empty window.
Compare
Copy the window and change the model, a parameter or the wording. Run all windows.
Constrain
Add a structured output schema or tools if your application needs them.
Save
Use Save as prompt to save the window as a new prompt, or as a new version of an existing chat prompt. Then move the production label when you are ready.