Model and conversation settings¶
Use the Technical settings to choose a model and decide how much conversation history the Agent can use. You can also connect knowledge and tools here. Change one setting at a time and check it with a test call.
Core settings¶
| Setting | What it changes | Starting approach |
|---|---|---|
| Model | Response quality, delay and supported controls | Select an available conversational model. Compare it with your actual call scenarios. |
| System prompt | Task, rules and fallback | Use the prompt guide. |
| Temperature | Response variation where supported | For an adjustable model, 0.3–0.5 is a useful test range for data collection. |
| Max completion tokens | Space available for the model response | Increase this value if responses are cut off. Use the Persona to ask for concise replies. |
| Conversation history limit | Amount of recent conversation | The default is 20. Include {conversation_history} in the prompt. |
| Knowledge Base | Retrieved factual context | Connect the relevant source and include {context}. |
| Voice response mode | Voice refinements and first-audio timing | Start with Quality first. |
Adjust these values based on your test calls.
Keep conversation context¶
To let the Agent use earlier details, set the history limit above zero and include {conversation_history} in the prompt. Add {conversation_fact_summary} to include saved facts.
Use the facts and history below to avoid asking for information twice.
Accept clear corrections. Ask for confirmation again if the callback number changes.
Stored facts:
{conversation_fact_summary}
Previous turns:
{conversation_history}
Keep {caller_phone_number} when confirming a callback number. Ask for a number if it is unavailable. Do not treat caller identification as verified identity.
Model-specific controls¶
After you choose a model, the editor shows its available settings. Some models have fixed values or offer fewer controls:
- GPT-5, GPT-5.6 and GPT-6 families have fixed temperature in the current configuration.
- GPT-5, GPT-5 Mini and GPT-5 Nano expose Minimal and Low reasoning effort. Minimal disables batched tool calls in one model response. Low permits them.
- GPT-5.6, supported GPT-6 models and GPT-6.1 Terra expose None and Low. GPT-6.1 Sol does not expose that configurable control.
- Controls differ for other variants. Use the options shown for the selected model and resolve validation errors before saving.
After changing models, retest complete spoken answers, tool selection, latency and cost. Do not copy a small output budget from another model blindly.
Add optional capabilities¶
Use search settings for Knowledge Base questions, tools for external actions and pre-processing for caller context. Explain in the prompt how to continue when information is missing.
Verify the configuration¶
Use Voice Debug Monitor during a test call to see the instructions and details the Agent receives. If a setting is missing, check your permissions.