Playground

The Playground is the safest place to improve quality. It lets you test changes without immediately affecting production.

What You Can Change In The Playground

The settings panel supports temporary overrides such as:

• model and OpenRouter variant
• temperature
• reasoning effort for supported models
• input token limit, or Auto for the provider context window
• conversation history summarization
• system prompt
• top K and similarity settings
• simulated language
• simulated channel
• interceptor testing
• retrieval settings

What Makes It Valuable

The Playground is more than a chat box. It also gives you:

• a debug panel for inspection
• persistent test sessions
• reset-to-production behavior
• a way to save validated settings back as production defaults

Temporary Overrides vs Saved Defaults

Use temporary overrides when you are exploring an idea or reproducing a bug.

Save defaults only when you are confident the change should become the new baseline for the agent. That keeps the Playground useful for experimentation instead of turning every test into a production edit.

Conversation history controls follow the same rule. A temporary override applies to that test session only; save it to make it the agent's baseline, or reset that control to Auto to inherit again.

What To Look For In The Debug Panel

The debug views matter because a good-looking answer can still come from the wrong process.

Check for things like:

• whether retrieval found the right chunks
• whether latency came from retrieval, generation, or tools
• whether an interceptor changed the path
• whether a simulated channel changed the answer style
• whether a model or threshold change improved the result for the right reason

Recommended Workflow

1. reproduce a real question from Conversations
2. test one change at a time
3. inspect the debug output, not just the final answer
4. compare against the current production default
5. only save changes after the improvement is clear and repeatable

Best Practices

• keep a short set of real benchmark questions for every major use case
• test both strong and weak queries, not just easy happy paths
• use simulated channel settings when answers differ by surface
• save defaults only after you have checked for regressions

Related Docs

Agent Settings

Tune AI, retrieval, escalation, and voice defaults for one agent.

Conversations

Review chats, reply as human, and inspect voice sessions.

Voice Lab

Test native voice sessions, local settings and backend tasks.