Overview
An Agent is the core entity in Vaani — it’s your AI-powered virtual phone assistant. Each agent is configured with:
- A system prompt that defines personality and behavior
- LLM, STT, and TTS providers for language processing and voice
- Custom functions (webhook-based tools the agent can invoke)
- Call behavior settings (max duration, silence handling, transfer config)
Creating an Agent
Navigate to Agents
In the dashboard sidebar, click Agents → Create Agent.
Configure Identity
| Setting | Description |
|---|
| Agent Name | Display name shown in dashboard and logs |
| Phone Number | Unique number assigned to this agent (auto-assigned or selected from provisioned numbers) |
| Agent Type | Category label for organizational purposes |
Choose AI Providers
Select from the available provider combinations:Language Model (LLM):| Provider | Models | Best For |
|---|
| OpenAI | gpt-4o, gpt-4o-mini, gpt-4o-realtime | Accuracy, tool calling, realtime voice |
| Groq | llama-3, mixtral | Speed and low latency |
| Provider | Models | Best For |
|---|
| Deepgram | nova-3 | English, multilingual, keyword boosting |
| Sarvam | Various | Indian languages (Hindi, Tamil, etc.) |
| Cartesia | Various | General purpose |
| Provider | Best For |
|---|
| OpenAI | Cost-effective, good quality |
| ElevenLabs | Most natural, premium voices |
| Deepgram | Lowest latency |
| Rime | Customizable voices |
| Cartesia | General purpose |
| Sarvam | Indian languages |
| Murf | Studio-quality |
| Inworld | Gaming/virtual character voices |
Write Your System Prompt
The system prompt defines your agent’s personality, knowledge, and rules:You are Sarah, a friendly appointments coordinator for Sunrise Dental Clinic.
Your responsibilities:
- Answer questions about our services (cleanings, fillings, root canals)
- Schedule, reschedule, and cancel appointments
- Collect patient information (name, phone, insurance)
Rules:
- Always be warm and professional
- If you can't answer a question, offer to transfer to a human agent
- Never provide medical advice
- Always confirm the appointment details before ending the call
Available appointment slots: Monday-Friday, 9 AM - 5 PM
Use double curly braces for dynamic variables that change per call: {{customer_name}}, {{appointment_date}}
Set the First Message
The first message is what the agent says when a call connects:Hello! Thank you for calling Sunrise Dental. My name is Sarah.
How can I help you today?
Configure Call Behavior
| Setting | Default | Description |
|---|
| Max Call Duration | 300s (5 min) | Hard time limit before auto-disconnect |
| Silence Timeout | 30s | Disconnect after this much silence |
| Ring Timeout | 30s | How long to ring on outbound calls |
| Agent Inactive Reminder | — | Prompt if agent is idle too long |
| Noise Cancellation | BVE | Background noise filtering model |
Add Custom Functions (Optional)
Custom functions let your agent take actions during calls by calling external webhooks:{
"name": "check_availability",
"description": "Check appointment availability for a given date",
"url": "https://api.your-clinic.com/availability",
"method": "POST",
"parameters": {
"date": {
"type": "string",
"description": "The date to check (YYYY-MM-DD format)"
}
},
"speak_during": "Let me check our availability for that date...",
"speak_after": "I've found the available slots."
}
Agent Settings Reference
LLM Settings
Voice Settings
Transfer Settings
Session Settings
| Setting | Type | Description |
|---|
LLM_provider | string | ”open ai” or “groq” |
LLM_model | string | Model identifier (e.g., “gpt-4o”) |
temperature | float | Response creativity (0.0 = deterministic, 2.0 = creative) |
max_tokens | integer | Maximum response length in tokens |
system_prompt | text | Agent instructions and personality |
first_message | text | Greeting when call connects |
| Setting | Type | Description |
|---|
STT_provider | string | Speech recognition provider |
STT_model | string | STT model identifier |
STT_language | string | Language code (e.g., “en”) |
stt_keywords | JSON | Words to boost recognition accuracy |
TTS_provider | string | Voice synthesis provider |
TTS_model | string | TTS model identifier |
TTS_voice | string | Voice ID or name |
TTS_speed | float | Speech rate multiplier |
| Setting | Type | Description |
|---|
transfer_phone_number | string | Target number for transfers |
warm_transfer_text | text | Briefing for the receiving party |
warm_transfer_retries | integer | Retry attempts for warm transfer |
| Setting | Type | Description |
|---|
max_call_duration | integer | Maximum call length (seconds) |
ring_timeout | integer | Outbound ring timeout (seconds) |
silence_timeout | integer | Auto-disconnect on silence (seconds) |
noise_cancellation_model | string | ”none” or “BVE” |
realtime_turn_detection_type | string | VAD algorithm |
realtime_turn_detection_threshold | float | Sensitivity (0.0 - 1.0) |
Knowledge Base
Upload documents to your agent for Retrieval-Augmented Generation (RAG):
- Go to your agent’s page → Knowledge Base tab
- Upload files (PDF, DOCX, TXT)
- Enable the RAG toggle
- The agent will now search uploaded documents during conversations and use relevant content in responses
Documents are indexed using LlamaIndex with OpenAI embeddings. The index is rebuilt when files are added or removed.
Best Practices
- Keep system prompts focused — agents perform better with clear, specific instructions
- Use dynamic variables for personalization:
{{customer_name}} and {{account_number}}
- Test with web calls first — use the dashboard’s Test Call feature before going live with phone calls
- Set reasonable timeouts — a 5-minute max duration prevents runaway calls
- Monitor cost per call — different providers have significantly different pricing
