This year at Microsoft Build 2023, we are announcing more generative AI features to preview in Microsoft Power Virtual Agents. Together, these features disrupt the traditional model of bot development and transform how conversational experiences are built today.
- New Generative Actions capabilities—Limited preview
- Generative Answers capabilities—Public preview
- Copilot in Power Virtual Agents—Generally available
- Unified bot building canvas—Generally available
- Azure custom language model integration—Public preview
- Instant generative bots for Microsoft Power Pages—Public preview
The new era of generative bot building
Bots typically do two things: provide users with answers (“How wide is this item?”) and complete actions by calling back-end APIs (“I’d like to make a payment”). Traditionally, both sets of experiences were built using hand-built dialogs. Bot builders had to anticipate every question or action a user might request, and then craft dialogs to handle each.
Now with generative AI and Power Virtual Agents, you simply point the bot to a knowledge source—a website, an internal SharePoint, or documents—and the bot uses generative answers to provide a rich multi-turn chat experience using the content.
But what about actions? Now you can point your bot to a collection of your tenant APIs, and when a request comes in that requires calling an action, the bot uses the new generative actions engine to intelligently call the right APIs to complete their requests.
Together, these two features transform the way bots are built today. Let’s take a closer look at each.
Conversations over your data with generative answers
We first announced our generative answers capabilities—which allow users to point to their public website and get answers from them—back in March 2023. Now, at Microsoft Build, we’re announcing the preview of the capability with numerous enhancements: You can now point to multiple knowledge sources—external sites, documents, internal SharePoint sites, or a combination of these—and promptly have rich, multi-turn conversations over that knowledge.
You can configure settings, customize the way the bot responds—and even build highly custom topics that can fetch data from your backend systems (for example, a billing or customer relationship management system)—and pass them into our new generative answers node to enable conversations over custom data.
Figure 1. Generative answers configuration and settings page with multiple data sources.
Figure 2. The new generative answers node within the canvas.
What’s more, if you have a Microsoft Azure OpenAI Service instance for your company data, you can hook it up with Power Virtual Agents to use with a few clicks, allowing your bots full access to those capabilities.
Try out generative answers now with a simple demo bot—just enter your company website and start chatting instantly. If you like what you see, want to extend with more knowledge sources, or author custom dialogs—you can sign up instantly for a free trial of Power Virtual Agents (for United States-based customers).
Turn conversation into action with generative actions
Generative answers can help answer user questions, but when actions and APIs need to be called, builders traditionally hand-craft dialogs around each. But that approach is fragile and time-consuming since it’s hard for a developer to anticipate every possible action a user might need to take.
Watch this video to see it in action
With our new generative actions capabilities—now available in Limited Preview—a bot maker can simply expose their tenant APIs, enterprise-specific Microsoft Power Automate flows, or other actions to the bot to use in conversation.
The bot understands the user’s request, looks through its library of actions, and identifies the ones that can fulfill the request. It then self-assembles and chains them together—and even generates questions for the user for any additional information it needs to complete the request.
The author can even see the bot’s rationale, watch it evaluate and decide
Figure 3. Test generative actions tracing and watch the chain log progress.
Taken together, generative answers and actions help usher in a complete transformation of bot building by helping bot authors be more productive. Bot authors no longer need to build and curate hundreds of dialogs for every question a user may ask—instead, they can just point the system to the right knowledge sources. There is no need to anticipate and hand-build dialogs for every possible common action that is needed—just point to the right set of connectors for existing APIs and the bot will help you with the rest. Between answering user questions by chatting over proprietary company information to dynamically assembling APIs to take action and complete user requests, Power Virtual Agents and generative AI have changed bot building for the better—saving bot authors’ time so they can focus on what’s most valuable for their organizations.
Power Virtual Agents—for professional developers and low-code developers alike
Even with the power of generative answers and actions, a bot maker may wish to supplement their bot with handcrafted dialogs—say, for a set of business-critical topics. Power Virtual Agents allows professional developers and low-code users to work together in the same experience to make this easy to do with the new advanced authoring canvas.
Low-code users can use the powerful graphical experience to build dialogs, use Microsoft Power Fx to manipulate data, build rich conversations with interactive adaptive cards, images, and videos, and create bots with Power Virtual Agents’ built-in universal large language models that require no custom model training. Copilot in Power Virtual Agents allows users to get started quickly—type in a sentence or two in natural language and a full multi-turn dialog will be created for you. You can even publish your bot to other low-code solutions including Microsoft Power Apps and Microsoft Power Pages.
Developers can now use all the capabilities of Azure Bot Framework natively in Power Virtual Agents—including a built-in code side-by-side editor, adding APIs or actions to the bot, trapping and triggering events, and manipulating data and variables—and even connect to Azure Cognitive Services (Preview) to create custom natural language models and more.
And low-code users and professional developers can collaborate in the same studio with our rich multi-authoring tools to build and maintain bots and add sophisticated actions.
Happy bot building!