Microsoft 365 Community Power Platform Connectors, 21 Sep 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Independent Publisher Connector - OpenAI connector on using AI for content creation and interactions

Description

In this 14-minute developer-focused demo, Robin Rosengrün shows how through his connector, users can access OpenAI’s Generative Pre-trained Transformer 3 (GPT-3), an autoregressive language model that uses deep learning to produce human-like text. To access GPT-3, you will need an OpenAI account. Connector exposes available GPT-3 List engines and responses to your questions (prompts) in any language. Familiarize yourself with OpenAI’s capabilities in Playground then build an instant cloud flow with an OpenAI step to access APIs via connector. This PnP Community demo is taken from the Microsoft 365 & Power Platform Development Community call recorded on September 1, 2022.

Demo Presenter
• Robin Rosengrün (EnBW) | @power_r2

Supporting materials
• Documentation - OpenAI (Independent Publisher) (Preview) | https://docs.microsoft.com/connectors/openaiip/
• Repo – OpenAI | https://github.com/microsoft/PowerPlatformConnectors/tree/dev/independent-publisher-connectors/OpenAI
• Connectors - Welcome to the Independent Publisher Connector Directory! | https://github.com/microsoft/PowerPlatformConnectors/tree/dev/independent-publisher-connectors/

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

A

Hi, my name is robin, and it's an absolute pleasure to be here today. I'm gonna talk about the openai gbt3 independent publisher, connector. I contributed this is available in all your tenants in all your power platform environments as long as it's not blocked by dlp, so you can set this up in a few minutes yourself. What I'm, showing today, let's take a first look at openai.

A

You may have probably heard about the company- does a lot of ai stuff uh yeah. What is news right now is one of the image generation um models, the dal e2 model, but that has no api, probably in the future, and so today we're going to talk about gpt 3 and what is gpt3.

A

It's short for generative pre-trained, transformer 3. It's an autoregressive language model that uses deep learning. I have no idea what all of this means. um I have no idea about ai, but it's so easy to use. We're gonna look into that right now, because it produces human like text. What can you do with it? um Some examples. Of course, this thing can do writing, and probably some of the news you are reading are already written by ai, probably by gpt 3, and this can do excellent fiction writing and is really creative.

A

This thing you see it in the top in the picture, is trained by 45 terabytes of data, so this thing basically knows the whole internet and at least the whole of wikipedia lots and lots of books. So it's really smart. It can bring up answers in a matter of seconds and and this thing speaks multiple languages and you don't set the language anywhere. You just start talking to it in language and it will respond in the according language and while english works, the best german is also a language.

A

This thing understands and also programming languages, and you will probably have heard um of this thing before the ideas tab in the canvas app editor- and this is actually powered by this model by gpt3. Of course, there's a little bit more magic going on in the background and it's uh trained a little bit further, so it can produce um powerfx.

A

But this thing can also already. The standard model can already produce um programming languages like html and others. So before we get into demo time, um I'm going to take you on a quick tour to get started and get your api key to use. My connector um setting up an account is done in a few minutes and these things actually cost money. But when you set up your account, um you get three months of free trial and eighteen eighteen dollars in credit, which is a lots to play around with this stuff and get started.

A

So you don't have to plug in credit card data. You can set this up in a few minutes and get going without paying a cent and where you should start is the playground, because this is basically what my connector can can do, and this will look very familiar if we look into into power automate in a second and first of all, um we're gonna plug in our first prompt that the string we are giving to the model is called a prompt.

A

um I was basically doing the thing that was in the hint text, but I um included a little bit more, so we are plugging in the prompt and we are getting a response back. This is yeah, um probably more creative than I would have been and makes absolute sense for ice cream shop, tagline, and you see how easy it was and basically we're doing the same thing with the api. There are some presets there down here that you can also change in the api. You should get familiar with them through the playground.

A

The most important is probably the engine that is in the background, and when you want to have human-like text, you should use the text davinci engine, which is the most the most expensive phone but um yeah. You can get familiar with the pricing and stuff and, as I said, you start with enough credits to play around a lot even with this model, and you should also check out the examples right here, because manufacturing your prompt is a science in itself. So let's look at one of my favorite examples: it's math!

A

Where is it math? The sarcastic chatbot right here? We open it in the playground, get back to the playground and get your prompt plugged in presets a little bit changed, but not too much and now we're gonna look at a little bit longer prompt right here. um It's separated in the first step.

A

This is the warm-up um where you just set the seam for the model, so it knows what you're expecting of it, and then you double down on that with this is called a few short learning approach, where you give on the spot a few examples um that you make sure you get the results, just as you want them, so we are giving here like four results or anything, and um then we can plug in it's already.

A

Everything set up we hit, submit and marv gives a sarcastic answer right now and lots of fun to play around with this example and one more thing and the model picks up really well when you use the colon, so you call in mouth colon that it's a dialogue between um you and the model itself, because we told the model that it's called math right now. So this is a technique that works really well, and this isn't programmed into the model. It's an ai. They just found out manufacturing prompts in this way works really really well.

A

So, let's jump into power automate first to see this um through the power platform connector, we will set up a new flow, probably not because it doesn't seem to work right now. So, let's give it one more try.

A

Now it works just an instant cloud flow which we will trigger manually and then add in a new step, open ai and once again this should be in your tenant as well, and this only has one end point the most important one, the completion endpoint- and this is the defaults that I plugged in which I would think, would showcase it a little from the get go.

A

We save it and test it for one time, so you see the same things you could you could put in in the playground and continue when you do it for the first time. Of course, you have to set up a connection enter your api key. You get from the open. Ai website really easy to find and we will run the flow and get our first response back through the api there. It is, and we ask what is the model's favorite animal and why and the favorite animal is an elephant because they end intelligent and loyal.

A

Also very large and impressive can weigh up to six or seven tons. Don't know if that is correct, but sounds impressive impressive. I want to point out one thing: um if you are not too familiar with response, objects or json objects, you're getting right here and we get this response object and then it has. This nested object choices which can have multiple records, and we generally want the first choice and then the text property of that that is yeah important.

A

We will look at the at the source code in our second example in a second- and this is the chatbot- it's a little bit like the like the playground like chatting with marv, but in a yeah in a canvas app. You can use on your smartphone really well, and this is actually in the powerapps samples repo.

A

I will plug a link in the end and, if you like design, then you should listen really carefully to christina's presentation, because this is from her from her new morphic design kit, which I used to to make this look more beautiful. I will plug a link at the end as well, so let's get started and ask the model what its favorite movie is, and the favorite movie is shawshank redemption. Indeed, a good one and let's ask for a short synopsis of the movie.

A

And we will get a perfect synopsis, it's a little bit longer. It's ended because I set the end condition to narrow. I think it would have wrote a little bit longer, but this probably does make a lot of sense and the rest would have made a lot of sense as well. I want to um add one more detail right here.

A

um This thing picks up context phenomenally, so I was just asking to write a short synopsis and of course, the model knows that I want a synopsis about the movie shawshank redemption, and how does this work? Let's look a little bit here, which prompt did we did we sent before um in each text we are sending again. We are including the whole conversation um for the prompt, so the model can also pick up on the context and give better answers right here and then we're gonna look at the second little demo app.

A

I have um equally lots of fun with that. It's my sample data generator, um which sample data are. We want to generate right here. We want five capitals of european cities, uh countries, sorry and we get paris, london, rome, madrid and berlin. This makes sense as well, so we were asking gpt 3 right here with our first call. This is a call to the google image search, um so no gpt 3 involved in here, and we get a really nice sample image.

A

We can scroll through some of them and we can ask gbt3 again for more information for a short description of the subject, and then we get a description about uh paris, london, rome and so on, and we can ask for another thing and we will ask for famous sites in the city of star and then internally, the star gets replaced with the city, and we will add this sublist right here. So for paris you get eiffel tower notre dame cathedral and so on, which seems to make a lot of sense and yeah.

A

I just love playing around with this and learning about different stuff through this app, so I'm having lots of fun and we are looking for the wrap up. We are looking for the api call we are making um through the canvas app right here.

A

We are calling the open ai, independent publisher, connector, the completion endpoint, which is the only one again available, plugging in the davinci model, and then we are setting it up just like we did it in the playground.

A

We are using a warm up and then a few short learning where I provide for four examples of um sample data that is generated and to process the data, it's important for me to have it separated by semicolons, which I also say in the warm-up and then what I said earlier. We wanted the first set of data from the choices object and from there we want the text. That is what we're usually after and then we are splitting the whole thing by semicolons, and that is where we're getting a nice collection.

A

We can work with, ah couldn't find any data, that's said, but here you can see that data is displayed through this collection. Then, of course, we are doing um exactly or similar things again with the description and the sublist and adding it to the collection, and that is it for my samples, I'm gonna plug the links in the in the chat and don't know how we proceed. Do we have time for some questions.

B

A

Usually the questions.

B

Get asked yeah the questions get asked in the chat. So definitely I believe there are some questions robin excellent. This is a super super cool, really love, how you've created it to be a little bit more fun and showed some business use cases for it really really flexible in terms of what you can do with this connector. So if you haven't checked it out definitely go check it out. Thank you for the links that are in the chat and by the way. Yes, this was chuck norris approved.

B

So just in case you were wondering so excellent job robin. Thank you.