Skip to main content
The SmartAI Block is the most powerful conversational node in the Verloop.io Voice AI recipe. Unlike standard blocks that follow a defined path, the SmartAI block uses Large Language Models (LLMs) to interpret user intent, handle complex logic, and engage in multi-turn conversations. This block is ideal for detailed decision-making scenarios, such as a debt collection agent that needs to negotiate payment plans based on a user’s specific financial situation.
Pro Tip: Choose the Right Block If you only need to ask a single question to capture a specific data point (e.g., “What is your date of birth?”), the Ask Block is better suited for the job. Use SmartAI for conversations that require reasoning or multiple steps.

​
Core Components

​
1. Prompt Section

This is the core “brain” of the agent for this specific block. Here, you define the persona, the goal, and the workflow the Voice AI Agent must follow. For the SmartAI Block, you should optimize for a detailed prompt that cleanly captures:
  • Use-cases and primary goals.
  • Edge cases (what to do if the user says something unexpected).
  • Boundary conditions (what the agent is not allowed to do).
Using Variables You can personalize the prompt using the {{variable_name}} format supported by the Recipe.
  • Example: “You are speaking to {{customer_name}} regarding a loan amount of {{debt_amount}} due on {{payment_date}}.”

​
2. Additional Instructions

This field is optional but highly recommended for defining behavioral guardrails. While the Prompt handles the “What,” the Additional Instructions handle the “How.”
  • Best Practice: Keep these pointwise and explicit.
  • Example: “Do not use slang. Maintain an empathetic tone. Never ask for the credit card CVV.”

​
Settings

The Settings section provides granular control over the AI Agent’s behavior, latency, and interaction style.

​
3.1 Retry Limits

“After how many tries bot should exit the block?” This setting prevents the agent from getting stuck in an infinite loop.
  • How it works: If the agent cannot determine an intent or collect the required info after the specified number of attempts, it will “hard exit” the block.
  • Where it goes: The flow will move immediately to the Fallback Node connection.
  • Use Case: Helpful when collecting optional information. If the user refuses to share it twice, the bot can move on rather than harassing the user.

​
3.2 Conversation History

Enable conversation history
  • Enabled (Default): The agent is aware of everything discussed previously in the call. This ensures a personalized and context-aware conversation.
  • Disabled: The agent treats this block as an isolated event.
  • When to Disable: Use this for mandatory disclosures or legal statements where previous context should not alter the agent’s strict adherence to the script.

​
3.3 Barge-in

Allow Barge-in for voice calls
  • Enabled: Users can interrupt the agent while it is speaking. This feels natural and humane.
  • Disabled: The agent will ignore user audio until it finishes speaking its current line.
  • Note: Even when disabled, user audio during the speech is recorded in the transcript.

​
3.4 Nudge

Behavior during silence
  • Enabled (Default): If there is an extended period of silence, the agent will repeat the question or prompt the user in a slightly altered way to keep the conversation moving.
  • Disabled: The agent will wait indefinitely (or until the call times out) if the user does not speak.

​
3.5 AnswerFlow (RAG)

AnswerFlow allows the agent to consult external documents and training materials to answer dynamic queries.
  • Document Tags: You can filter which documents the agent accesses by selecting specific tags (e.g., refund_policy, credit_card). If left blank, the entire library is used.
  • Response Formatting: You can define how the answer should be delivered (e.g., “Keep it under 2 sentences,” “Use a bulleted list”).
Latency Impact Enabling AnswerFlow requires the system to search and retrieve documents. This will add a small amount of latency to the Agent’s response time.

​
3.6 Speech Normalization

This converts raw text data (like “100mg” or “14/02/2024”) into natural spoken text (like “one hundred milligrams” or “the fourteenth of February, twenty twenty-four”).
Only add normalizations that are strictly required for this specific block. Each active normalization rule adds processing time, leading to higher latency.

​
3.7 Intent and Variable Mapping

This section defines the “Success” exit conditions for the block.
  • Intents: Define the specific user intents that, when detected, will cause the agent to exit this block and move to the next step in the recipe.
  • Variable Extraction: You can configure the LLM to extract specific data points (e.g., payment_amount) upon exit.
Performance Note Extracting variables via the SmartAI block increases latency.
  • Recommendation: Avoid extraction unless the data is immediately required for logic in the very next block.
  • Alternative: If the data is only needed for analytics, use Post-Call Insights to extract it after the call concludes.

​
3.8 Intent Passing

Intent to be passed to the next block Specify a variable name here. The intent detected within this SmartAI block will be stored in this variable, allowing you to use it for routing logic further down the workflow.

​
Fallback Node

Every SmartAI block must have a Fallback Node connection. The workflow routes here if:
  1. The Retry Count (Section 3.1) is exhausted.
  2. The AI Agent encounters a system error or critical failure.
Ensure this node connects to a graceful error message or a human handoff to prevent call drops.