Test in the playground
Try changes without saving, compare two models and inspect exactly what a reply was given.
On this page
The Playground tab is where you talk to your agent as a customer would, before any customer does. You can try new settings without saving them, compare two models side by side, and see exactly what each reply was based on.
Talk to the agent
- Open the agent and select the Playground tab.
- Type a question as a customer would ask it. Enter sends; Shift+Enter adds a new line.
- Read the reply. Use 👍 / 👎 to mark good and bad answers as you go.
Playground conversations are kept apart from real ones. They never appear in the Inbox or in Analytics. Replies still use credits, as they do anywhere.
Try changes without saving
Above the conversation you can change the model, temperature and instructions. As soon as they differ from the agent's saved settings, a Not saved badge appears:
- Replies in the playground now use your new settings. The live agent keeps its own until you save.
- Save to agent makes them the agent's settings in one click.
- Reset throws your changes away.
This makes a safe loop for improving an agent:
- 1Ask a real question
- 2Change one setting
- 3Ask again
- 4Save to agentwhen it is better
Compare two models
Select Compare models to split the playground into two panes. Each pane has its own model, temperature and instructions. One message goes to both, so you see the two answers to the same question side by side. Stop comparing returns to one pane.
Use this to decide whether a more expensive model is worth its credits.
Inspect a reply
Every reply shows how long it took, the tokens it used and what it cost. Open Inspect reply to see what the model was given at that moment:
- Knowledge: every passage found for the question, each with its similarity score, including the ones that scored too low to be used.
- System prompt: the instructions exactly as sent.
- Temperature and the conversation history.
This is how you find out why an agent answered as it did:
| What you see | What it usually means | What to do |
|---|---|---|
| The right passage is there, but the answer is wrong | The instructions or the model | Make the instructions more specific, or try another model |
| The right passage is missing, or scored too low | The knowledge | Add or reword a source, or add a Q&A pair for that question |
| "This agent has no knowledge yet" | Nothing to search | Add knowledge |
Related
- Knowledge sourcesTeach an agent from files, websites, text and questions with their answers.
- Create your first agentName it, pick a model, write its instructions and set its temperature.
- TroubleshootingWhat to check when the widget does not show, an agent cannot answer, training fails or a channel is silent.
Something not working as described here? See Troubleshooting.