Call Stages

The Ultravox API’s Call Stages functionality allows you to create dynamic, multi-stage conversations. Stages enable more complex and nuanced agent interactions, giving you fine-grained control over the conversation flow.

Each stage can have a new system prompt, a different set of tools, a new voice, an updated conversation history, and more.

Understanding Call Stages

Call Stages (“Stages”) provide a way to segment a conversation into distinct phases, each with its own system prompt and potentially different parameters. This enables interactions that can adapt and change focus as the conversation progresses.

Key points to understand about Stages:

Dynamic System Prompts → Stages allow you to give granular system prompts to the model as the conversation progresses.

Flexibility → You have full control to determine when and how you want the conversation to progress to the next stage.

Thoughtful Design → Implementing stages requires careful planning and consideration of the conversation structure. Consider how to handle stage transitions and test thoroughly to ensure a natural flow to the conversation.

Maintain Context → Think about how the agent will maintain context about the user between stages if you need to ensure a coherent conversation.

Advanced Feature

Stages require planning and careful implementation. Stages are likely not required for simple use cases.

Creating and Managing Stages

To implement Call Stages in your Ultravox application, follow these steps:

1. Plan Your Stages

Determine the different phases of your conversation and what prompts or parameters should change at each stage.

2. Implement a Stage Change Tool

Create a custom tool that will trigger stage changes when called. This tool should:

• Respond with a new-stage response type. This creates the new stage. How you send the response depends on the tool type:
  • For server/HTTP tools, set the X-Ultravox-Response-Type header to new-stage.
  • For client tools, set responseType="new-stage" on your ClientToolResult object.
• Provide the updated parameters (e.g., system prompt, tools, voice) for the new stage in the response body.

Unless overridden, stages inherit all properties of the existing call. See Stages Call Properties for the list of call properties that can be changed.

3. Configure Stage Transitions

• Prompt the agent to use the stage change tool at appropriate points in the conversation.
• Ensure the stage change tool is part of selectedTools when creating the call as well as during new stages (if needed).
• Update your system prompt as needed to instruct the agent on when/how to use the stage change tool.

Things to Remember

• New stages inherit all properties from the previous stage unless explicitly overridden.
• Refer to Stages Call Properties to understand which call properties can be changed as part of a new stage.
• Test your stage transitions thoroughly to ensure the conversation flows naturally.

Example Stage Change Implementation

Here’s a basic example of how to implement a new call stage.

First, we create a tool that is responsible for changing stages:

Basic stage change tool

function changeStage(requestBody) {
  const responseBody = {
    systemPrompt: "...", // new prompt
    ... // other properties
  };

  return {
    body: responseBody,
    headers: {
      'X-Ultravox-Response-Type': 'new-stage'
    }
  };
}

We also need to ensure that we have instructed our agent to use the tool and that we add the tool to our selectedTools during the creation of the call.

Updating our systemPrompt and selectedTools

// Instruct the agent on how to use the stage management tool
// Add the tool to selectedTools
{
  systemPrompt: "You are a helpful assistant...you have access to a tool called changeStage...",
  ...
  selectedTools: [
    {
      "temporaryTool": {
        "modelToolName": "changeStage",
        "description": ...,
        "dynamicParameters": [...],
      }
    }
  ]
}

Inheritance

New stages inherit all properties from the previous stage. You can selectively overwrite properties as needed when defining a new stage.

See Stages Call Properties for more.

Ultravox API Implications

If you are not using stages for a call, retrieving calls or call messages via the API (e.g. GET /api/calls) works as expected.

However, if you are using call stages then you most likely want to use the stage-centric API endpoints to get stage-specific settings, messages, etc.

Use GET /api/calls/{call_id}/stages to get all the stages for a given call.

Ultravox API	Description	Stage-Centric Equivalent	Description
/calls/{call_id}	Get a call	/calls/{call_id}/stages/{call_stage_id}	Get the call stage
/calls/{call_id}/messages	Get messages for a call	/calls/{call_id}/stages/{call_stage_id}/messages	Get message for the stage
/calls/{call_id}/tools	Get tools for a call	/calls/{call_id}/stages/{call_stage_id}/tools	Get tools for the stage

Stages Call Properties

Tip

The schema used for a Stages response body is a subset of the request body schema used when creating a new call. The response body must contain the new values for any properties you want to change in the new stage.

Unless overridden, stages inherit all properties of the existing call.

Here is the list of all call properties that can and cannot be changed during a new stage:

property	change with new stage?
systemPrompt	Yes
temperature	Yes
voice	Yes
languageHint	Yes
initialMessages	Yes
selectedTools	Yes
firstSpeaker	No
model	No
joinTimeout	No
maxDuration	No
timeExceededMessage	No
inactivityMessages	No
medium	No
initiator	No
recordingEnabled	No

Use Cases for Call Stages

Call Stages are particularly useful for complex conversational flows. Here are some example scenarios:

Data Gathering → Scenarios where the agent needs to collect a lot of data. Examples: job applications, medical intake forms, applying for a mortgage.

Here are potential stages for a Mortgage Application:

• Stage 1: Greeting and basic information gathering
• Stage 2: Financial assessment
• Stage 3: Property evaluation
• Stage 4: Presentation of loan options
• Stage 5: Hand-off to loan officer

Switching Contexts → Scenarios where the agent needs to navigate different contexts. Examples: customer support escalation, triaging IT issues.

Let’s consider what the potential stages might be for Customer Support:

• Stage 1: Initial greeting and problem identification
• Stage 2: Troubleshooting
• Stage 3: Resolution or escalation (to another stage or to a human support agent)

Conclusion

Call Stages in the Ultravox API give you the ability to create adaptive conversations for more complex scenarios like data gathering or switching contexts. By allowing granular control over system prompts and conversation parameters at different stages, you can create more dynamic and context-aware interactions.