Skip to main content
The LLM Playground is a chat-style interface for testing an agent’s language model behavior in isolation. You type messages as the caller, and the agent responds as text. There is no voice, no transcription, and no telephony involved. Use it for prompt development, function call testing, and logic validation before moving to audio or end-to-end testing.

​
What the LLM Playground Tests

  • Prompt instruction following
  • Response quality and tone
  • Handling of unexpected or off-topic inputs
  • Function and tool call behavior (does the agent call the right function with the right arguments?)
  • Context window management across a multi-turn conversation
  • Variable substitution in the prompt

​
What It Does Not Test

  • Voice quality or pronunciation (no TTS involved)
  • Transcription accuracy (no ASR involved)
  • Response latency on a real call
  • Telephony carrier behavior
  • Audio interruption and barge-in handling
A prompt that works correctly in the LLM Playground may still have issues in a real call due to transcription errors or audio timing. Always follow up with an audio test or full test call before deploying to production.

​
Open the LLM Playground

  1. Navigate to your agent in the DialNexa dashboard.
  2. Click the LLM Playground tab in the agent detail view.
  3. The playground opens with the agent’s current draft configuration loaded.
The playground always uses the current draft version of the agent. To test a published version, publish your changes first, then open the playground.

​
Use the Playground

​
Send a Message

Type a message in the input field at the bottom and press Enter or click Send. The agent’s response appears in the conversation thread.

​
Set Variables

If your agent prompt references variables (e.g., {{customer_name}}), set their values before starting the conversation:
  1. Click Variables in the playground toolbar.
  2. Enter key-value pairs for each variable referenced in the prompt.
  3. Click Apply. The conversation resets and variables are injected into the prompt.

​
Review Function Calls

When the agent triggers a function or tool call, the playground displays a structured function call block showing:
  • The function name called
  • The arguments passed
  • The response returned (mocked or live, depending on your tool configuration)
This lets you verify that the agent calls the correct function with the expected arguments under different conversation paths.

​
Simulate Function Responses

For tools that make external HTTP calls, you can configure mock responses in the playground:
  1. Click Tools in the toolbar.
  2. Find the tool you want to mock.
  3. Toggle Mock Response and enter the JSON response body to return.
This lets you test how the agent handles different tool responses (success, error, empty result) without making live API calls.

​
Reset the Conversation

Click Clear in the toolbar to reset the conversation to the beginning. The agent’s memory of the conversation is cleared. Variables remain set.

​
Example Workflow

Testing a support agent that looks up order status:
  1. Set variables: customer_name = "Alex", account_id = "acct_7842".
  2. Send: "Hi, I want to check on my order."
  3. Verify the agent’s greeting uses {{customer_name}} correctly.
  4. Send: "My order number is 9988."
  5. Verify the agent calls the lookup_order function with order_id = "9988" and account_id = "acct_7842".
  6. Review the mock function response handling.
  7. Send an off-script message: "Can you also book me a flight?"
  8. Verify the agent declines gracefully and stays on topic.

​
Limitations

LimitationDetails
No voice or audioTTS and ASR are not exercised. Pronunciation issues will not appear here.
No latency measurementResponse speed in the playground does not reflect real-call latency, which depends on the full stack including transcription and audio streaming.
No barge-in simulationThe playground does not simulate a caller interrupting the agent mid-response.
Draft version onlyThe playground loads the current draft. To test a specific published version, you must check out or recreate that version’s configuration.