Creating a Custom Chatbot

CS123, Intro to AI

Contents

Copilot Studio

Full details on creating a custom GPT chatbot using Microsoft Copilot Studio are provided in the Microsoft documentation for Copilot Studio.

There are essentially two ways a chatbot created in Copilot Studio can generate answers:

• Topics: This approach is based on using keywords in the user's query to trigger topics that have branching conversation flows that follow a decision tree (remember this from symbolic logic?)

• Generative: An LLM is used with RAG to generate responses to the user's queries.

Tips

Generative AI Responses

• To get your copilot to use generative AI to respond, go to Settings, Generative AI, and click the radio button labeled "Generative (preview)"

• In order to get the copilot to use my web site, in "Overview" I had to turn off "Allow the AI to use its own general knowledge (preview)".

Greeting

The initial greeting comes from the topic: "conversation start". You can edit it in "conversation flow".

Authentication

Click on the "Settings" button, "Security", "Authentication". These are the options

• No authentication: This is the reccomended option for our class assignment.

• Authenticate with Microsoft

  If you want to use OneDrive as a knowledge source, you have to use this option, which uses Microsoft Entra ID¹, but, then you won't be able to publish to a web site!

• Authenticate manually

Knowledge Sources

Click on the "Add Knowledge" button, there will be two options, choose "Featured". The sources shown here are sources that don not require a special enterprise data connection:

• Public web sites (must be indexed by Bing)

  • You can see if the site you want to use is indexed by Bing by using the site: search operator. For example, in the Bing search bar type: site:yourdomain.com This will show all the pages from your site that have been indexed.

• Files

• OneDrive folders—Requires Microsoft Entra ID (Teams and Office 365)

Index for Searching

All knowledge source depend on the source for searching. This means the files need to have their content indexed. It happens different ways in different knowledge sources:

• Web site: your site needs to be indexed by Bing. This can take days (or more) and depends on getting your site ready according to the Bing Webmaster requirements.

• Files: will be made searchable by Dataverse.

• OneDrive: Will be indexed by OneDrive almost immediately after upload.

Document Types for Files

• Markdown and html
  These files could be in one of three knowledge sources: web site, files, or OneDrive.

  • In a web site: just html will be indexed, citations (links in the response) open as a web page.

  • In files: (stored in Dataverse) both will be indexed, citations open as text only.

  • In OneDrive: neither will be indexed

• PDF and Word

  • PDFs and Word docs get indexed everywhere

  • PDF citations open as documents, Word docs get downloaded instead of opening

• Excel and .csv

  • Both can be searched and understood when uploaded as files.

Publishing to a Web Site

You can publish your chatbot in many different ways as can be seen on the "Channels" page for your chatbot. To put your chatbot on a web page you have two options:

• Custom website: This option will provide you with code that you can put on your own web site to embed the chatbot on a page.

• Demo website: This option will create a web page that is hosted by Microsoft and that contains your chatbot. You can share the link with anyone so they can use your chatbot.

References

Microsoft Copilot Studio Documentation

Intro to AI lecture notes by Brian Bird, written in 2024, are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Note: Microsoft Copilot with GPT-4 was used to draft parts of these notes.