Skip to main content

Retell AI Complete Reference

Overview​

Retell AI is an API for building human-like conversational voice AI agents with response times averaging 800ms.

Voice Agent Types​

  • Single Prompt Agents: Best for straightforward conversations
  • Multi-Prompt Agents: Multiple specialized prompts for different phases
  • Conversation Flow Agents: Visual flow editor with conditional logic

Custom Functions​

{
  "name": "book_appointment",
  "description": "Books an appointment in the calendar system",
  "url": "https://your-api.com/book",
  "method": "POST",
  "parameters": {
    "type": "object",
    "properties": {
      "date": {"type": "string"},
      "time": {"type": "string"},
      "name": {"type": "string"}
    }
  }
}

Webhooks​

Event Types:

  • call_started: When call begins
  • call_ended: When call completes
  • call_analyzed: Post-call analysis available
  • function_called: Custom function executed

Pricing (2025)​

  • Voice synthesis: $0.07-$0.08/minute
  • LLM usage: $0.03-$0.08/minute
  • Telephony: $0.01-$0.03/minute
  • Total: $0.13-$0.31/minute

Prompt Engineering Template​

You are [role description]. Your goal is to [main objective].

CONVERSATION FLOW:
1. Greeting: [specific greeting]
2. Qualification: [what to ask]
3. Action: [what to do with info]
4. Closing: [how to end]

RULES:
- [Specific rule 1]
- [Specific rule 2]

TONE: [Professional/Casual/Empathetic]

WHEN TO CALL FUNCTIONS:
- [Function name]: [When to use it]

Common Use Cases​

Appointment Booking​

  1. Agent checks calendar availability via function
  2. Collects customer preferences
  3. Confirms slot via function call
  4. Sends confirmation SMS
  5. Updates CRM with appointment details

Lead Qualification​

  1. Ask qualifying questions per script
  2. Calculate lead score based on responses
  3. Route qualified leads via webhook
  4. Send lead data to CRM
  5. Trigger follow-up workflow

Troubleshooting​

Agent not calling functions:

  • Verify function description is clear
  • Check parameter schema is correct
  • Test in LLM playground first

High latency:

  • Use turbo models when available
  • Optimize function response times
  • Reduce knowledge base size