Skip to main content

​
Optimization Guide

To get the most accurate results from your Voice Agents, follow these prompt engineering best practices.

​
Dos and Don’ts

The “Description” field is effectively a prompt for the LLM.
  • Bad: “Check interest.”
  • Good: “Did the user explicitly state they want to buy the insurance policy? Mark as True only if they said ‘Yes’ or ‘Agree’.”
When using Enum (Select from List), ensure your options cover all scenarios.
  • Recommended: Always include an option like "Unclear" or "Other" so the AI isn’t forced to make a wrong guess if the user was vague.
Keep descriptions under 300 characters. If you need complex logic, split it into two separate parameters (e.g., one for “Interest” and one for “Reason”).

​
Handling Latency

Analysis is evoked immediately but takes time to process.
  • Expectation: Data is usually ready within seconds but may take up to 5 minutes post-call.
  • Workflow: Do not block your immediate post-call logic (like sending an SMS) on the analysis result. Use the Webhook to trigger a separate follow-up flow once the data arrives.

​
Regional Nuances

For Voice AI Agents for Arabic or Indian regions, the transcription might contain mixed languages.
  • Tip: If you are analyzing a Hinglish call, you don’t need to translate the description into Hindi. The AI understands the context of the transcript even if the prompt is in English.
  • Tip: For Arabic, ensure your description asks for the intent rather than looking for specific keywords, as dialects (Khaleeji vs. Levantine) use different words for the same action.