Mastering Generative Answers in Copilot Studio

Getting Started with Co-Pilot Studio

• Co-Pilot Studio connects to websites, SharePoint, and custom data sources for AI-based chat.

• The experience can be improved; guidance is provided to enhance user satisfaction.

• Traditional chatbots require predefined topics, limiting awareness of user questions.

Generative Answers and Data Sources

• Generative answers utilize various data sources to provide comprehensive responses.

• Currently using GPT 3.5 turbo model hosted on Azure OpenAI service.

• The service is included in the Co-Pilot Studio license costing about $200/month.

Retrieval Augmented Generation (RAG)

• RAG supplements a generic model with business data without extensive training.

• Users can ground the co-pilot experience by pointing it to public websites or documents.

• This method is resource-efficient compared to training a custom model.

Conversation Context and Boosting Topics

• The system retains context from the last 10 conversation turns for better interaction.

• Conversation boosting topics enable generative answers as fallback options when needed.

• Users can build chatbots without predefined topics or mix them with specific intents.

Limitations and Advanced Options

• Generative answers can be used within specific topics or as overall fallback solutions.

• Limitations exist; advanced tools are available for more complex setups at higher costs.

Understanding Document Upload Limitations

• Maximum of four sites can be pointed to for document uploads, with a size limit of 3 MB per file.

• Dataverse search indexes documents to help find context but cannot provide direct links to uploaded documents.

• SharePoint or OneDrive offers a better experience for document management compared to direct uploads.

Using Public Data for Co-Pilot

• Example used is The Better Health Channel website, which provides health information.

• The bot retrieves results from the specified website without needing predefined topics.

• URL depth limitation is two levels; however, it can retrieve results from deeper levels.

Generative AI Capabilities

• Users can request content in different formats, such as simplifying explanations for children.

• Content moderation settings (high, medium, low) affect answer accuracy and creativity.

• Low moderation may yield creative but less accurate responses; high moderation ensures more reliable answers.

Common Issues with Bot Responses

• Users may encounter "I can't help with that" errors due to various reasons including setup time.

• If no results are found, it could be due to the data source lacking relevant information.

Understanding Course Information Retrieval

• Discusses issues with retrieving course information from a university website due to different top-level domains.

• Emphasizes the importance of website structure for successful information retrieval; adding relevant domains can yield results.

• Notes that if a website has multiple top-level domains, it may complicate search results.

Document Uploading and Citation Challenges

• Explains uploading policy documents into co-pilot for quick answers about parental leave.

• Highlights limitations in formatting when documents are uploaded, affecting user experience.

• Suggests that while document uploads are simple, they may not be practical for real-life applications.

Comparing SharePoint and OneDrive Experiences

• Contrasts experiences between direct document uploads and using SharePoint for better user interaction.

• Points out that linking to documents in SharePoint enhances accessibility compared to plain text responses.

• Mentions the need for authentication when accessing documents on SharePoint.

Enhancing Search Experience with Microsoft 365

• Introduces Microsoft 365 semantic search as a tool for improved indexing and search capabilities within co-pilot.

• References a video explaining how this feature works to enhance user experience across various applications.

How to Use Custom Instructions for Generative AI

Understanding Custom Instructions

• Custom instructions allow users to set a persona for the AI, enhancing interaction quality.

• Providing meta prompts influences the AI's responses, leading to better user experiences.

• Example: Acting as an HR assistant with specific parameters improves response tone and relevance.

Tone and Context in Responses

• Adding warmth and positivity (e.g., emojis) can change the tone of responses significantly.

• Adjusting instructions based on context (e.g., manager vs. staff interactions) is crucial for appropriateness.

• Removing casual elements like emojis may yield more professional responses in serious contexts.

Length and Brevity of Responses

• Custom instructions can specify response length, aiding clarity and conciseness.

• Users should experiment with character limits to achieve desired brevity in answers.

• Effective prompting is essential for generating concise and relevant outputs from the AI.

Importance of User Prompting

• Educating users on effective prompting enhances overall experience with generative AI tools.

• Clear guidelines help users formulate questions that lead to useful responses from the AI.

• Specific phrasing in prompts can improve the accuracy of information retrieved by the AI.

Enhancing Output Quality

• Providing detailed prompts allows the AI to generate more targeted responses related to policies or issues.

• Using structured requests (e.g., bullet points for meeting preparation) yields clearer outputs from the AI.

Understanding Retrieval Augmented Generation

Key Insights

• Emphasizes the importance of simplifying complex information for better understanding, akin to explaining it to a child.

• Highlights the capability of handling multi-turn conversations and the need for improvement in rewording requests.

• Discusses challenges in getting concise scripts for meetings, indicating limitations in current capabilities.

Performance Improvement Plans

Documenting Progress

• Focuses on documenting milestones within performance improvement plans, leveraging previous conversation threads effectively.

• Stresses understanding data sources and use cases to optimize retrieval augmented generation outcomes.