Parameter deep dive¶
Tuning guide
Parameters are for fine-tuning behavior that is already understood. Use them deliberately: one change at a time, with the same test scenario before and after.
Do not use advanced settings to hide an unclear prompt.
Parameters control how the agent behaves, speaks, and uses resources. They are powerful, but they should not be used as a substitute for a clear prompt.
Change one thing at a time
If you change the prompt, model, temperature, and voice settings at once, you will not know which change caused the result.
Agent and model settings¶
| Setting | What it affects | Recommended approach |
|---|---|---|
| Model | Reasoning quality, speed, and cost | Use the default or recommended model unless you have a specific reason to change it. |
| System prompt | Agent behavior and boundaries | Fix prompt problems here before changing advanced settings. |
| Temperature | How varied or creative answers are | Lower for strict, predictable workflows. Higher for more flexible conversations. |
| Max tokens | Maximum answer length | Keep high enough for useful answers, but not so high that the agent becomes verbose. |
| Knowledge base | Company-specific facts | Use for FAQs, policies, products, and documents. Keep behavior instructions in the prompt. |
| Tools | Actions the agent can perform | Add only tools the agent truly needs. Describe when to use them. |
Temperature¶
Temperature changes how much variation the model may use.
| Use case | Suggested direction |
|---|---|
| Compliance-sensitive answers | Lower temperature |
| Routing decisions | Lower temperature |
| Data collection | Lower temperature |
| Friendly open conversation | Medium temperature |
| Creative wording | Higher temperature |
If the agent ignores instructions, do not immediately lower the temperature. First check whether the instruction is clear and testable.
Max tokens¶
Max tokens limit the length of generated responses.
Use shorter limits when:
- the agent should answer quickly,
- callers need concise voice responses,
- the workflow is form-like.
Use higher limits when:
- the agent summarizes long conversations,
- the agent explains complex topics,
- the agent uses long knowledge-base answers.
Prompt variables¶
Some prompt editors show available variables. These values are injected at runtime.
Rules for variables:
- Use only variables shown in the editor.
- Keep variable names exactly as shown.
- Do not invent variables.
- Test the prompt after adding variables.
- Include a fallback instruction for missing values.
Example:
If the caller's phone number is available, use it for follow-up.
If it is not available, ask the caller for the best callback number.
Voice settings¶
Voice settings affect how the voice agent sounds. Availability depends on the selected voice.
| Setting | What it affects | Use carefully when |
|---|---|---|
| Voice | Speaker identity and language/accent | The brand needs a specific tone or language. |
| Expressiveness | Emotional variation in speech | The agent sounds too flat or too dramatic. |
| Speech tempo | Pace and focus | Callers need slower explanations or faster confirmations. |
| Word focus | Vocabulary variety | The voice sounds repetitive or too unpredictable. |
| Speaking style | Emotional style, if supported | You need a specific tone such as cheerful or calm. |
| Accent override | Language/accent output | The voice uses the wrong accent for your callers. |
Voice settings do not fix unclear behavior
If the agent says the wrong thing, fix the prompt or knowledge base first. Voice settings only affect how it sounds.
A safe tuning workflow¶
- Start with defaults.
- Test real conversations.
- Identify one concrete problem.
- Decide whether it is a prompt, knowledge, tool, routing, or voice issue.
- Change one setting.
- Test the same scenarios again.
- Keep the change only if it improves the result consistently.