Intents

Introduction to the Intents Page

For more information about using Voiceflow's new intent classification using LLMs feature, see documentation here.

Voiceflow's Intent CMS is a powerful tool that helps you define and manage the intents for your AI agent. An intent represents what a user means or intends to do when they interact with your application.

Intent Table Overview

• Name: Displays the name of the intent. This should be descriptive and indicative of the user's goal when triggering the intent.
• Description: A brief summary that gives context to the intent, explaining its purpose within the conversational flow.
• Confidence: An indicator that shows how confidently the AI can recognize and match the user's utterance to the intent.

• Used by: Displays the workflows and components that are utilizing this intent, providing insight into its dependencies and connections.

• Last Editor: Indicates the last person (user) who edited the intent. This is helpful for teams working collaboratively.
• Updated: Shows when the intent was last modified, providing a timeline of changes.

The CMS allows you to sort the intents to better organize your view. Here are common sorting options you may have:

1. Alphabetical: Sort intents by name in alphabetical order, either A-Z or Z-A.
2. Confidence Score: Arrange the intents by their confidence score, either from highest to lowest or vice versa.
3. Last Editor: Sort by the last editor, which can be useful when tracking changes by specific team members.
4. Updated: Order the intents based on the most recently updated to the least recent.

Creating an Intent

1. In the Intent CMS, click on the 'New intent' button in the upper right area of the screen.
2. A modal will appear prompting you to enter details for the new intent.
3. Enter a name for the intent in the 'Name' field—choose a name that clearly represents the user's goal, such as "BookAppointment".
4. Provide a description in the 'Description' text box, which explains the intent’s purpose, like "This intent is for users looking to book an appointment."
5. Under 'Utterances,' type in examples of what users might say to trigger this intent, e.g., "I want to schedule a meeting," "Can I book a time for a call?"
6. If the intent requires specific information, click the '+' icon next to 'Required entities' to add entities like "date," "time," or "service type." 
7. Enter the reprompt text to the required entity. This text should guide the user to give the specific information that is missing. For instance, if the entity is 'date', the reprompt could be, "What date are you interested in?"
8. Once all information is entered, click 'Create intent' to save the new intent to your CMS.

Editing Intent Details

1. In the intent editor, select an existing utterance, required entity or description to edit it.
2. Add new utterances to improve recognition by adding new utterances or clicking 'Generate' to have an LLM generate example utterances on your behalf.
3. Remove irrelevant utterances or required entities by clicking the '-' to remove them from the list.
4. Add or edit your entity reprompts in the required entities section.
5. Make sure to update the description if the intent's purpose has evolved over time.

Deleting an Intent

Single Intent

1. To delete a single intent, select the intent you wish to remove, and in the editor view click the '...' (more options) button, and select 'Delete' from the dropdown menu.
2. Confirm the deletion when prompted to remove the intent from your CMS.

Bulk Deleting Intents

1. For bulk actions, select multiple intents by selecting the checkboxes next to the intent names.
2. Once selected, on the bulk action toolbar above the table, click the 'Delete' button to initiate the bulk deletion process.
3. Confirm the bulk deletion to remove all selected intents simultaneously.

Preview

Previewing your intent classification method is an essential step in ensuring the accuracy and performance of your agent's ability to understand your users queries. This feature simulates real-world interactions, allowing you to assess and refine the agent’s intent recognition.

1. Access the Preview: Locate the 'Preview' button positioned in the header of the Intent CMS page.

2. Input Your Utterance: In the 'Utterance' field of the preview modal, enter a question or query that a user might say. This should be representative of the actual queries you expect once your agent is live.

3. Send the Query: After typing your utterance, click the 'Send' button. The system will process the query using the intent classification method you've selected.

4. Review the Response: The result will be displayed in the classification section. There are two versions of this depending on which method is configured:

   • Classify intents using NLU: The NLU classification section will show up to 10 intents triggered from the utterance, sorted by confidence score. The first result will be the selected result from the NLU.

   • Classify intents using LLM: An additional section called LLM classification will appear above the NLU classified intents which returns the intent triggered by the LLM. 

   • NOTE: Only intents that are used in your agent will be seen in the results.For example, you will need to add intents to a supported step (see Usage 
     section below) for it to be considered in preview.

5. Feedback: You can optionally provide feedback by selecting the thumbs up or thumbs down button in the results section. This data will be used, in aggregate, to improve Voiceflow's intent classification services. 

   • If you select thumbs down, there will be an option to select which intent was the intended target of your utterance.
     
6. Refine and Iterate: Based on the preview results, you may find it necessary to adjust your training data or tweak the model settings to improve the accuracy of your model. You can select the 'Re-use last utterance' button to resend the previous utterance again with the updated settings. 

Tips for Effective Previewing:

• Test with Variety: Use a wide range of utterances to ensure robustness in the models intent classification accuracy.
• Test with Different Settings: View the previewing process as iterative. The goal is to continuously enhance the models ability to respond accurately. 
  • Preview-Specific Settings: Changes made in Preview are temporary and will not impact the agents overall settings. For example, if you wanted to try different AI models when using LLM classification, you can do so easily in the preview without updating the model used by your agent.
  • Overriding Defaults: To test different behaviours, you can adjust the AI model, temperature, and prompt wrapper which will override the global intent classification settings. You can see how many settings have been changed, or reset back the settings back to default by selecting the link at the top of the settings modal.

• • Permanent Changes: If satisfied with the preview results, you must manually update these preferences in the Intent Settings to apply them globally.

•  

Usage

Intents that you create can be utilized in your designs in the:

• Intent Step
• Button Step
• Choice Step

Note: Intents and Entities created in the Intent CMS are available 
globally within the assistant.

Best Practices for Intents

• Be specific: Create intents that are focused on specific tasks, avoiding broad or ambiguous intents that can confuse the AI.
• Use representative utterances: Provide a variety of examples that cover different ways users might express the same intent.
• Capture variability: Include synonyms, slang, and common misspellings in your utterances to account for the diverse ways users may communicate.
• Categorize logically: Group intents in a way that makes sense for your application, such as by function or by the part of the conversation they belong to.
• Maintain consistency: Follow a consistent naming convention and structure for your intents to make them easily identifiable.
• Update regularly: Add new utterances based on real user interactions to keep the AI’s understanding up to date.
• Balance quantity and quality: While having more utterances can improve accuracy, ensure that each utterance is a quality example that adds value.
• Analyze confidence levels: This helps determine how well each intent is understood by the system.
• Perform routine testing: Testing to ensure the intents are capturing user queries accurately.