Microsoft 365 Community Microsoft Teams, 15 Feb 2023

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Create a conversational bot in Microsoft Teams using OpenAI

Description

In this 15-minute developer-focused demo, Lee Ford steps viewers through the process - user sends message from Teams client to the Azure Bot ID defined in Teams app manifest, Azure Bot sends message to Bot code that is running in an Azure Function (doesn’t have to be an Azure Function), Bot sends the text from message sent from Teams to OpenAI API and answer/response is sent back to Teams (in Adaptive Cards).

This PnP Community demo is taken from the Microsoft 365 & Power Platform Development Community call recorded on January 19, 2023.

Demo Presenter
• Lee Ford (Symity) | @lee_ford

Community Call Conversation
• aka.ms/Jan19-Demo1

Supporting materials
• Sample - Teams OpenAI Conversation Bot | aka.ms/bot-openai
• Org – Open AI | https://openai.com/about/
• Documentation – GPT-3 model | https://beta.openai.com/docs/models/gpt-3

Learn more
• Microsoft 365 Unified Sample gallery - https://aka.ms/m365/samples
• Microsoft 365 Platform Community in YouTube - https://aka.ms/m365/videos
• Microsoft 365 Platform Community - http://aka.ms/m365/community

👋 Join to our weekly Microsoft 365 & Power Platform Community calls to learn more. These are PUBLIC calls for you to get access on the latest details from Microsoft and to see awesome demos from community on these technologies - Download recurrent invites from https://aka.ms/community/calls

A

A

With open API, open, AI, not open, API, that's completely different and the idea is we're kind of go through the sample and show you how it all hangs together. So, as I said, I'm a solution architect for a company called simity based in the UK, and hopefully it will. Let me pick my there. We go so a bit about myself, I'm based in the UK in Birmingham uh to be exact.

A

um You can follow me on Twitter I've got GitHub and obviously you can follow my blog, which I do blog very occasionally, I'm, a Microsoft MVP in M365 development and I, mainly work in Microsoft Technology, so n365 Azure and the like. Obviously I'd like to write code for a living. That's why I'm on this call and when health permits I do like to go for a run but uh yeah there we go so so what is a conversational bot?

A

Exactly and that's the term that I've just sort of coined for this session um and we'll go over? What it's not first, which is it's not a q, a box? So we're not asking it a question. Expecting an answer like what is two plus two or you know: what's the weather today, it's it's more kind of um uh advanced than that. um It's also not a uh fixed or predetermined path, bar or sometimes called a water flow. A waterfall dialogue where you know you've got a predetermined path that you know.

A

If you answer option A, you know for the first question, it's then going to present three.

A

um You know uh other questions down the road that are predetermined because you chose option A at the at the beginning of the of the of the conversation. That's again not what we're we're talking about conversational bot. What we're talking about is a bot that you can have a conversation with without any outcome or paths. So you know you can put whatever you uh you want as a question or as a task, and you don't necessarily know what you're going to get back, because it's not sort of being coded or predetermined. What's going to happen.

A

So um what who or what is open, AI um so I did think about using chat GPT to ask that question, but instead this is actually what openai call themselves so they're a research and deployment company- um and you know they're, trying to effectively uh help AI, be you know, adopted uh out uh sort of out there in the world, obviously to benefit all of humanity. So let's see how that goes over the years. So what can you do with open AI?

A

So um there are different products and there's probably ones that I'm not going to cover here, but the sort of the three uh main ones that we're going to talk about here are a Dali 2, which has been around a little while now um I think there's some samples of this running um already with uh sort of M365 uh products.

A

But the idea is that you can generate an image from a description, so you could say you know, um draw a you know a panda living on the moon or you know something really random and the idea is. It will try and generate that base on the description and the more descriptive you are generally the better. The uh the generated image is um so that's one uh another one is gpt3, that's what we're going to talk about today in a few minutes, and this one is um sort of I, say natural language conversation.

A

So the idea is that it's not predetermined. You can kind of have a back and forth with uh with the uh with the gpt3 uh product, and it's going to kind of in theory, um yeah lead to a lead to a good outcome and the final one is Chap GPT, which everyone probably already knows what it is. It's quite a Hot Topic at the moment. um Obviously, that's currently in preview, there's no apis for it.

A

um So you know you have to use it through there that web page, you can kind of have session tokens and all that sort of stuff, but we're not gonna I'm, not gonna cover that today, um uh so that is um kind of an enhanced version or a trained version of gpt3 or specifically GPT 3.5. That's been tweaked to sort of better leader conversation than sort of the standard gpt3 product. So a quick word on GPT in a bit more detail.

A

um It's not to be she confused with chat, uh chubby GPS, not to be confused with qpt3. They are kind of um Ones based on the other, but they are separate so um today we're not going to cover chat GPT, as I mentioned, there's no real API that I can integrate uh with yet so it doesn't make sense to have a session on it really. um So, um instead we're going to talk about qpt3 as they are supported within a uh an AI API, um and we can go over that.

A

So uh what is gpt3 it understands and generates natural language, so it's not necessarily just generating the natural language as in you know, write me a poem about you know whatever, and it also needs to understand what we're asking it. So it could be a simple. You know what is two plus two or it could be. You know uh yeah two paragraphs asking it to do something: they're different ends of the um of the spectrum, but ultimately they're both uh natural language.

A

So this is used using different models, so in in gpt3 there is more than one model and we'll cover that now um we have four models um and each model has like an iteration of it. So the top one uh DaVinci zero, zero three. So that's the third iteration of the uh the model um that is generally the most advanced one. So that's where you're going to get the best results.

A

um Ultimately, there is a a higher cost associated with that one and then obviously we have Ada, which is the most sort of basic I guess, uh and that's generally used for smaller tasks or more simple tasks that um don't involve particularly complicated uh scenarios.

A

So we'll quickly go over the solution that we have it's a very simple solution. We have a in this case, A team's client, that's connected to an Azure bot registration or an Azure bot, and then that connects up to a HTTP uh request trigger uh in an Azure function. um Today the demo I will run locally, but it could be hosted in Azure. So it's a very simple stack and that connects to an open, AI API to uh to allow the uh the sort of the AI to kind of form a response.

A

So the general flow is team's. Client will send a message to the uh to the Azure bot. The Azure bot will send it to the uh the Azure function. The Azure function will then send that text that's been sent from teams to the API. The API respond with the with an answer, and that then goes down to the back then to the client. So it's a very straightforward scenario: I'm losing my voice, so I do apologize.

A

um Hopefully you can hear me okay, um so let's just quickly go to the demo. um So let me quickly share with you in live.

A

um So, as you can see here, I've got a conversation, I'm asking you what its favorite number is. uh It's told me it's 17, but literally five minutes ago told me it was eight, and now it's seven, so uh you know very silly conversation, but it kind of shows you that there is, you know a bit of variance and and a bit of sort of playfulness with the AI. So I can now ask it a different question.

A

So I can say uh what is your favorite drink I have no idea what it's going to say, but we'll find out what percent. That is something to think about it. Iced Green, Tea, okay, I'm, not even sure what that is, but I'm sure it's delicious, so again, yeah, um okay. uh What else do you like and again we can ask you another question and it's going to think about it. uh Okay, I was more about the drink, but okay, you can be a bit more uh open than that. Okay, so there we go.

A

That's what they like, um so we'll quickly cover the uh sort of the the Azure part of it now, which is uh generally in in the in the. In this scenario, I've only got the the Azure box. Obviously if it was running in Azure, you might have the Azure functions in here as well, um but I will show that in in sort of log running locally to show you how it all works.

A

So as far as a uh configuration goes, we've literally got a bot profile. uh You can give it a name, upload, a logo um and then from a configuration point of view, as I said, as it's running locally on my machine I've got engrock to uh sort of tunnel the uh the bot back through to my local machine and that's about it from a configuration point of view in here. It's very straightforward um from a channel perspective, obviously we're using teams, but you can use other channels.

A

um The sample does actually work with other channels. Obviously it's a team's focused uh session, but there's nothing to stop you using the code uh in other in other ways. Okay, so, let's quickly go into uh showing you that so if I run it in web chat, so this is obviously not in teams. I still get the uh the Bots and I can say: uh do you like cheese? Okay, let's see what they say: it's election? Okay, okay, is your favorite okay and again.

A

So this is a different conversation to the one in teams it's using the same code. uh Okay, that's not quite right, but okay, that's AI for you, um okay, so that is it from the configuration point of view, I think uh in Azure, so I'll quickly. Flip over to the game, just that we don't have too much time so um I might just make that slightly larger for everyone to see.

A

So this is the sort of the main entry point into the into the Azure function, so um we're effectively creating a uh an open, AI bot uh instance and we are um authenticating. So if I just show you the code here for the authentication part we are using um in this case, because we're running locally we're using an app password, but the code is configured in a way using the the cloud adapter that we could absolutely use MSI.

A

So if this is running in Azure, we don't need to use flight secrets and whatnot, so um we're effectively receiving the message, we're creating an instance of the or getting an instance off the bot and then we're sending it to the uh the bot to run uh the the code. So if we look at the uh the code for the bot, it's actually quite straightforward, um we have two activity handlers, so I'll cover the first one. I saw the second one first, which is a very straightforward one.

A

So when a member is added to the chat so like we were showing you in um the the web chat, we send a worker message, pretty straightforward, a bit of code here we have an Adaptive card. That's obviously uh got the welcome message in and then we go through and we we send the attachment as an activity.

A

Just say: hey, welcome, ask me any question whatever and then the main part of the code is in here, and this is where we're using an open, AI API key to get the um uh be able to call the open, AI API and we're using the official open, AI module which we're importing up here and passing it a prompt, uh we're, also processing it the model that we want to use.

A

So if you remember before, there was four different models and we are- and we can I'll show you how we can figure that in a minute, but it's pretty straightforward and we're also configuring how many tokens so the more tokens you provide it generally, the better response or the better uh or essential potential for a response that you get back so maybe we're getting not great results today. In this particular demo, um with with the answers of giving, if you give it more tokens, maybe you'll get a better better answer back.

A

So once we've got the uh uh we create the completion. We um we sort of wait for the response. Once once the response comes back, we choose the first choice because generally there's only one choice that comes back from a um from an answer. We take the text from that and we pipe that into an Adaptive card. So we've got the uh the answer card, which I'll show you in a minute um and we effectively inject the answer into the card.

A

Once that's been ejected, we then send the adapter card back to the uh the team's client as an activity. So we attach the car to the activity and send it from a configuration point of view. In teams uh it's pretty straightforward. We set up a an act which uh sorry app manifest we put in where we've got the app ID. That's the app ID of the. um If I quickly go back to the, uh and here we've got the the app ID that's generated when we create the Azure bot.

A

So we would specify that as the uh the app ID and we just need to specify in a couple of uh a couple of places, so I use it in the ID of the app, but you could use a different ID there, but mainly it needs to be in here um and you can just obviously specify the ID and what Scopes the um the bottom uh with inside of that's it for the Manifest and then finally I'm going to change these. So don't worry!

A

You don't need to screenshot these, but these are you know where we have the ID uh yeah uh up, ID the attendant ID. uh Obviously the app password like I say in in production, you're gonna be using msis and no one should be doing this. Obviously because it's running locally on my machine for the demo um I'm, obviously uh using a password, the Open Eye model, again I'm using the DaVinci 003. You know in your scenario.

A

You may not need something as advanced as that you could use an ADA or the uh the Babbage or something like that where you want something a little bit: um yeah less expensive.

A

um So each um you're calling the the API is not costing you as much uh and then obviously, we've got the API key, which is required to record the API. um Anyone can sign up for the open, AI API account um and you that you get like I think it's about twenty dollars worth of credit. Something like that that you can. You can spend within three months, I think so you know. If you want to try this today, you don't have to worry about register for an account and waiting to hear back.

A

You can kind of get started today and then finally, I'll just quickly show you the uh the Adaptive card. So we are very straightforward format. Here, we've literally got two um text blocks um and this is the welcome card. It's just saying ask me a question and then the answer card is even simpler.

A

We literally have some large text with the answer which gets injected into the uh into the card and I think that's it from a code point of view, so I'm just gonna go back to the the deck, and that thing is to say thank you for listening. I do have a um uh sort of the code provided, I've put it into the uh the dev sample EMP repo, so obviously uh yeah welcome to have a look and try it out for yourself. Thank you really great stuff, Lee excellent.

A

Thank you for sharing love the love, the fun that you can have there with the bot uh and all of that.