Generative AI Activities

Overview

BotStudio provides Generative AI activities that enhance chatbot responses through AI-driven content generation. These activities enable chat summaries, AI-generated replies, and classification of user queries using AI models.

Note: Generative AI functionalities may require an extended license. Contact Puzzel for activation details.

Screenshot Placeholder: Insert an image of the Generative AI activities section in BotStudio.

Generative AI Activities

BotStudio includes the following Generative AI activities:

1. Generate Chat Summary – Creates a concise summary of an ongoing conversation.
2. Generate AI Reply – Automatically generates chatbot responses based on conversation history.
3. Generative AI Classification – Uses AI to categorize user queries for better flow management.

Screenshot Placeholder: Insert an image of the Generative AI activity selection menu.

Generate Chat Summary

The Generate Chat Summary activity provides a short recap of the conversation, making it easier for agents to understand chat context before taking over.

Enhancing Summaries with Additional Data

The summary can be improved by extracting specific information using:

• AI Data Extraction: Allows AI to detect key details from the chat (e.g., "Topic" or "Username").
• BotScript Values: Uses pre-existing chatbot variables or named entities.

Example:

• The AI extracts whether the issue is related to cancellation, payment, or login.
• It retrieves the user’s phone number if detected during the conversation.

Screenshot Placeholder: Insert an image showing a configured chat summary with extracted fields.

Generate AI Reply

This activity enables the chatbot to create AI-generated responses based on:

• Chat transcript
• Selected context sources

Choosing a Context for the AI Reply

Administrators can specify the data source for AI-generated responses:

1. Search Engine – Uses a SupSearch engine and a predefined query.
2. Article – Pulls responses from selected SupSearch articles.
3. Manual – Manually enters context from a website or document.

Additional context can be added via BotScript variables to refine the AI’s response.

Screenshot Placeholder: Insert an image of AI Reply settings with context selection options.

Response Control & Formatting

• Fallback Handling: If AI cannot generate a useful reply, the chatbot redirects to the Fallback Node.
• SupSearch Feedback: If enabled, article clicks automatically improve search result accuracy.

Screenshot Placeholder: Insert an image showing AI-generated response formatting options.

AI Safeguards and Guards

AI Safeguards

The Safeguard feature allows administrators to control how AI responds by enforcing specific wording rules:

• Ensure responses contain or exclude certain words.
• Apply no restrictions (default behavior).

AI Guards and Their Functions

AI Guards help determine whether an AI-generated response should be sent by evaluating different aspects of the generated message:

1. Word Similarity Guard
   • Ensures AI-generated responses use similar wording to reference articles.
   • Example: If the chatbot response is answering a query about "password reset," but the answer contains unrelated terminology, it may be rejected.
2. Semantic Similarity Guard
   • Evaluates whether AI understands context rather than just matching words.
   • Example: The phrase "Can I book a flight?" and "Is it possible to schedule a flight?" mean the same thing, even though the words are different. This guard ensures such responses are considered valid.
3. Response Relevance Guard
   1. Assesses whether the AI's response is relevant to the specific user question.
   2. Example: If a user asks about pricing but the AI responds with information about cancellation policies, this guard would block the response.

4. Article Groundedness Guard

   • Measures how well the AI response aligns with the reference knowledge base, without requiring identical wording.
   • Example: If an AI-generated response says, "Your refund will be processed in 5-7 business days," but the article states "Refunds typically take one week," this guard ensures alignment in meaning.

    

Note: Article Similarity Guard and Response Relevance Guard may incur additional costs.

Screenshot Placeholder: Insert an image displaying AI Guards with threshold settings.

Sending Generated Replies

Users can configure:

1. Threshold for AI-generated responses – Controls confidence level required before sending a response.
2. Article Link Display – Optionally shows links to source articles.
3. Fallback Handling – If AI fails to generate a response, the chatbot defaults to article links or fallback paths.

Screenshot Placeholder: Insert an image of reply settings with threshold adjustments.

Generative AI Classification

This feature enables AI-powered classification of user queries, allowing the chatbot to determine intent dynamically.

Creating a Classification Activity

1. Navigate to the Node Editor.
2. Add Generative AI Classification as an activity.
3. Click Add Category to create classification labels.
4. Provide example phrases for each category.
5. Assign a Confidence Score for classification accuracy.

Screenshot Placeholder: Insert an image of category creation in Generative AI Classification.

Example Use Case

• The chatbot classifies queries related to Puzzel support.
• AI assigns a category (e.g., Puzzel_Identifier).
• The chatbot routes users to relevant nodes based on classification.

Screenshot Placeholder: Insert an image showing category-based node routing using Generative AI Classification.

Improving AI Classification

To enhance classification accuracy:

• Add more examples per category.
• Create a fallback category to handle unmatched queries.
• Use multiple classification activities for granular intent detection.

Screenshot Placeholder: Insert an image of expanded AI classification examples.

Best Practices for Generative AI Activities

• Use chat summaries for smoother agent transitions.
• Implement AI safeguards to control chatbot messaging.
• Set AI guard thresholds to ensure high-quality responses.
• Regularly update classification examples to improve intent recognition.
• Monitor AI-generated replies for accuracy and optimize thresholds accordingly.