Boost conversation for PVA – When should you use it?

Shows the Create screen of Power Virtual Agent. The name of the Bot is entered as Ben Bot and we will wire up the Conversation Booster to

ChatGPT and OpenAI has taken the world by storm. Now almost everyone is talking about AI and especially generative AI. First DALL-E shows us we can generate images and art just by prompts then ChatGPT shows we can generate entire blogpost by doing the same. I find these AI models thoroughly impressive but what I find even more so is how fast Microsoft and other tech companies are incorporating these models into their products. Today I want to focus on one small piece of these announcement and give it my thoughts by asking myself the following question: When and how should you use Boost Conversation for Power Virtual Agents?

Let’s start off with a brief description of what Boost Conversation is for Power Virtual Agents. Boost Conversation allows you to wire up your PVA to a website of your choosing. Meaning it will read the content of that website and index it. And then tries to answer the questions a user ask to your Chatbot with data found on that website. For this it uses Azure OpenAI services paired up with Bing to access and index that website. So instead of using ChatGPT 3.5 to try to answer the question with the knowledge of the entire internet it will narrow it down. Is this a brilliant decision or a flaw by the product team?

The Pro’s of narrowing ChatGPT down

Now I think it’s quite smart to integrate Boost Conversation this way. Just think about how we are using Chatbots currently. We are using Virtual Agents to help us perform tasks in a conversational manner. This will make the interaction flow between user and bot. This however has the downside that getting the correct information or interaction across somewhat longer. Thus we use Chatbots for relatively short interactions for specific intent such as filling a service request. I will not use a companies Virtual Agent as a creativity starter for my next Talk at an event 😉

Shows a ChatGTP question and answer. The question asks to help write a session submission to a conference. The answer from ChatGPT is very detailed and helpful as a starting point.
Help me write a session (click to enlarge)

We will look at my website for example. What if I want to add a Virtual Agent to my site? The purpose of my blog for you, the reader, is to share my knowledge. For me, my blog is a lot about gathering my thoughts and extending my research in certain areas. And as a side-benefit, I am also promoting my persona. So with this in mind, if I create a PVA for my blog, I want it to help you find the content you need from my site. And if this was a commercial website, I definitely do not want you to redirect away from my site. This is one of the reasons why I like that PVA is narrowing down Azure Open AI.

Shows the Create screen of Power Virtual Agent. The name of the Bot is entered as Ben Bot and we will wire up the Conversation Booster to
Creation of your Boosted Bot

So now that we’re talking about this example, let’s build upon it. I now have wired up PVA to What I am immediately curious about is the Bot Content Moderation option. This will help in scaling the creativity of the model. More creativity could mean less accurate answers. See below for a quick test I did with the question: “Who is Ben den Blanken?”.

Shows what happens to the question "Who is Ben?" when you set the moderation to high. The answer is quite strict and short.
High Moderation Response
Shows what happens to the question "Who is Ben?" when you set the moderation to medium. The answer is a bit more elaborate.
Medium Moderation Response
Shows what happens to the question "Who is Ben?" when you set the moderation to low. The answer is a lot more lengthier.
Low Moderation Response

Now is a good time to talk about the when

Okay, I now have a PVA ready to go for my website right? It will answer any questions the reader might have about my content. But does it actually give the information that I want? In fact, Microsoft proposes to use Conversation Booster as your final fall back before escalating to an agent. The order Power Virtual Agent will use to answer the question is as follows:

  1. Authored Topics.
  2. FAQ – Scripted responses.
  3. Conversation Booster generated responses.
  4. Escalation to a live agent.

Continuing on reverse engineering these steps. Take the example of my question above “Who is Ben?” or “What is Ben den Blanken specialty?”.

Shows a chat conversation where the user asks What is Ben den Blankens specialty. The answer is generic and not necessarily helpful.
Is this really my Specialty?

I am not too thrilled about the answers generated here. Granted, I should be updating my about page so Bing can do a better job summarizing who I am. But I think this is a perfect example when you would rather use an Authored Topic. I can present my information the way I want in a card as follows:

Shows a chat conversation where the user asks Who is Ben. The answer is now authored by the Bot builder and shows exactly what I want.
Ben as a Card

Now that does look a lot more like me and how I want to present myself. This is also what you need to think about when you are building your Power Virtual Agent. What message do I want to give, how do I want to present myself?

Where does it fit?

The use case for Boost Conversations is definitely around the more general Bots that try to answer the first few questions before handing off to a Live Agent. It gives substantial body to your Chatbot and will probably save you a lot of time instead of building all the topics or FAQs.

I still want to play around with it a bit more. I want to know how I can go from a Boost Conversations answer to escalating to an agent. There also is functionality to gather feedback on the answers given by Boost Conversations. This will help decide where you as a Conversation Designer want to add authored topics.

All in all it’s an impressive first release. Not only the Large Language Model from OpenAI integrated into PVA, but also all these little features surrounding Boost Conversations. I hope this blog helps you to find the place where this feature fits best. Let me know what you think!

Leave a Reply

Your email address will not be published. Required fields are marked *