Introduction to Azure OpenAI and Architecture Patterns

[Music] thank you [Music] the channel will be talking about a lot of different stuff but we needed to get started today with open AI so I'm starting to get a lot of customers and I think a lot of the information I can cover in a video instead of having one-on-one conversations with lots of customers um so with that let's just go ahead and get started I think everybody's heard of openai they've played with open AI what we're going to talk about today is Introduction to open AI the models a little bit and we're going to talk about some possible architecture patterns as well as give you some demos so with that we'll go ahead and get started I have a GitHub repo which will be linked at the bottom of the video so you'll be able to open this repo and get started if you uh and this is about Azure open AI specifically not open AI the company or open AI the public the out the Microsoft out facing so you can imagine that as your open AI would actually be an open AI model deployment in your subscription and we'll talk about what that actually means throughout the video but the first thing you're going to want to do is get access to it so on the GitHub link you'll actually see that the register for open open AI access this will give you access to basically GPT 3.5 right now it takes uh less than a day in many cases that may change in the future uh because we're pretty much letting a lot of folks get access to it so they can do some tryouts and keep in mind that this is still as of today uh mid-june we're still in basically a public preview so before you go to production you might want to think carefully about you know what if we Yank A feature because that has happened in the past uh gpt4 access once you have access to this you can apply for gtp4 as of mid-june they've actually they're blocking any new customers from coming on board just due to demand we're trying to keep up we'll go through some of these other links a little bit later the Azure AI content safety Azure pricing page we'll talk a little bit about that why it matters related to the model and then the most important thing that you're going to need to talk to your security and legal team about is data privacy and security inside of azure open AI uh because that's one thing a lot of customers get tripped up on they're like yeah we don't want you using our data to train your next model on and that's an absolutely Fair Point and we don't for what it's worth but we'll talk about that and you've got the links here to that you can go through as well so to start with we're actually going to go to the gpt3 models page on the Azure documentation you can search for this pretty easily but what I wanted to talk about is the fact that when you get open AI enabled in your subscription it's not one model that you have access to it's all of these models that you can get access to especially with the gpt3 models obviously when you get access to gpt4 there are some other models that uh that are be available to you the gpt4 and the the 432k which actually has a more tokens and we'll talk about that in just a second but what I wanted to point out on this one was that you'll see that the uh the main models we're talking about are the Ada the Babbage the Curie and The DaVinci okay so those are going to be the models that you're typically going to use to get you started with whatever open AI feature you want text summarization document summarization respond to an email help me write an email things like that um the max request tokens for DaVinci if you look at this is going to be for DaVinci 3 Tex DaVinci 3 is going to be 4 000 tokens uh for the rest of these it's going to be around 2 000 tokens so then the question becomes what is a token well basically a token is how a word gets divided up uh one way you could think about it is uh is a as is as a syllable even though that's not exactly exactly accurate another way to think about it is each word is approximately 0.75 tokens on average so if you've got a sentence that's got you know 10 words in it you can pretty much figure out that you know multiply that out and you'll figure out that you know you could have uh you know 14 12 13 14 15 tokens in that sentence depending on the size of the word so that's one way to think about it and there are some tools out there and you can actually Implement uh in your code to count the number of tokens and you can actually send it and get a token request back as well to see how large it's going to be but that's something to keep in mind is uh you can't send in to DaVinci a 5 000 page document and expect it to come back with a summary because it's too many tokens and there are some architecture patterns we're going to talk about that will allow you to do that now there's another thing we need to discuss here as well and the best way to demonstrate it is on the pricing page so if we look at the pricing details for these like I said before Ada Babbage Curie and da Vinci are the core models that you're going to be using for text analytics and that's a b c d so you can kind of get them in order and remember them that way it's kind of a mnemonic you know ABCD mnemonic that you can kind of remember which order they come in and which capability they come in now you can see that per 1000 tokens DaVinci is going to cost you about 2 cents per call so for 2000 tokens four cents per Call Etc and you'll notice the Curie Babbage and Ada are less expensive the reason for that is the capability of the model DaVinci is the most robust model curious slightly below that Babbage is slightly below Curie and Ada is slightly below Babbage so one thing I would think about and I've been talking to customers about is when you start your POC and I wouldn't just throw you know an open AI model into production with no testing but I would definitely do a proof of concept in such a way that maybe you start with DaVinci and then you see how Da Vinci is behaving with you and your users which you know your your POC should include a few actual users so you can kind of get an idea of the behavior of the model and the behavior of the users because there's going to be some prompt engineering even when you ask it a question that we're going to talk about later where you need to figure out it's like what's the proper way to communicate with a model so but one thing you can do is come out start with da Vinci in the POC and then maybe do some a b testing against Da Vinci and Curie and see if Curie will provide enough information if it does a good enough job because you can save yourself quite a bit of money or even Babbage or Ada in some cases and then obviously we have the code the code once which if you've looked at GitHub copilot that's an example of the code Cushman and the code DaVinci and obviously the code DaVinci is relatively expensive as is the code Cushman so if you decide to create your own co-pilot for software engineering you might want to be cognizant of the prices of that and then chat GPT which is going to be basically the interactive chat session that we're all used to seeing right now and there are some use cases for chat GPT as well as for text DaVinci most of what we're going to be using today is going to be text DaVinci that we're actually going to go through so we'll get started with that next all right let's go back to our GitHub repo for just a second we've got our our links out here so one of the things I've got a useful repos in here open AI links and resources which we talked about already I've also got a section called useful repos and this will grow over time as I find more uh more content more folks that have published content that I think would be interesting or useful for you so we'll actually go through that a little bit today but there are a few that are out here that I like I've deployed uh all of these at some point except for the openai.net samples I haven't deployed that one yet and I've actually been through these Jupiter notebook samples as well so one easy way to get started is to actually come out here and start looking at these notebooks because the penalty for getting started is very low have python have Jupiter installed and have an open AI account in Azure and you can go ahead and get started so that's going to be relatively simple for you to uh to work on but we'll go ahead and get started let's start by looking at some of these uh Microsoft Azure open AI samples okay with that let's go ahead and get started with uh one of our repos just go down and look at our Jupiter notebook we're actually going to start out with the Azure open AI examples this is probably one of my favorites uh so we'll jump into that what we're going to be going through today if you drill into that we're going to go to examples uh demo you'll notice there's several outer you can play with we're going to start with the demo uh the demo IPython notebook so you can take a look at this in your GitHub repo and just clone the repo just like you do with everything else but let's go ahead and get started and see what that looks like so I've already got this set up and this is typically what you're going to do to get your Jupiter notebook to communicate with Azure open AI notice we're using the open AI package and we've got the service type API type the base the version the key and the model that we're using that's actually set up in Azure open AI and you'll get your own URL and you'll get your own key from that so it's not a big deal but let's go and get started uh so create a list of 10 locations with no internet white sand beaches no cars and the island is accessible only by canoe so what we're going to start with is we've got a temperature of two we've got some things we got on here the engine model which we said is DaVinci The Prompt which is this I think we set up up here and a temperature of two and Max tokens is the number of tokens to return temperature of two is kind of important because two gives it basically it's telling it you've got a lot of flexibility in what you can return so when you do two you could get a different answer every single time but in this case we're just asking open AI it's like hey based on your own model training what you know what can we actually do with this if you were to set this to say zero for instance and run that again which you'll get back is the same answer every single time so there's no flexibility in the answer so the temperature is going to be important when you're searching your own documents because you don't want it to be very creative when it's searching your own documents doing a summary and giving a return you want the answer to be pretty much the same every single time so that's something to think about so if we run the zero again we'll actually see that we're going to get the same I assume the answers back again and if we go back to you know and it ranges from zero to two but if we go back to go back to a temperature of two you'll see that basically we could get a random response every single time and that opens up the world for hallucinations as well so be careful it's like you don't want if you're searching your own documents you don't want to be doing that other things you can do with open AIS this is a fun one we actually are passing in a table as a string into the stock prompt so you can see here we have basically stock information for a day or for a couple of months actually and you can ask it a question to take the stock prompt which is the data that this table is and then concatenate on what did this stock close at on 120. so you can come in here and give it you know an answer uh oops I can run the whole thing so we'll run that and when it'll come back with an answer we'll see that it closed at 84.72 on 120. and if we go look at 120 we'll see that it did in fact close at 84.72 and this is another one where you want to be really careful with your temperature because you may want to set this to zero because you don't want it being imaginative in its responses so if we set that to two for instance let's see if it comes back with the correct answer of 84.72 luckily it did so but yeah to keep that in keep that in mind when you're working with temperature you may want it to be very specific in what it comes back with what was the average price and this is another one where you're doing mathematics with this you're going to want to do some testing to make sure that it's coming back with the right answer every single time the other one I don't have the code prompt set up right now but we you can also do a code prompt in here uh where you can actually pass in you know the following you know basically give it the uh stock table that we were just using and then load the variable stock prop as a CSV file and we're giving it instructions on having it create a pandas uh script for us or excuse me a python script for us using pandas so that we can actually do some stuff on it uh this is what it came back with when the original Arthur wrote the wrote this I don't have the code set up in my environment right now because it would instead of being Da Vinci it would actually be da Vinci Code it'd be a different model that we would be using we'd be not not being using the normal chat GPT but like I said when you get your uh when you get your subscription enabled for open AI you'll actually see that there's a lot of models that you have available to you that you can use in any way generated the code and then you can actually pass in take a look at that uh this is where this is where uh basically Azure open AI starts to shine and this is where people start lighting up typically is what can I do with it uh in this case we have a contract prompt so let's go ahead and start with that and load it up um and we're gonna I'll go ahead and get a lot running for you uh but yeah so in this case definitions I agree to acknowledge and represent the following blah blah blah uh what happens if I miss a payment for instance so this is a loan contract uh and it will come back with a response and ask you know what if you miss a contract it'll tell you you know the lender and other loan participants May report this to information to credit bureaus that sounds fair uh what can the lender do with my information and that'll actually give you a summary of what from that contract what can the lender do with your information so this is a good one where you can actually have it write a summary uh or ask it questions against the data set and it will come back with an answer and once again be careful on your temperature because that's where you're going to get some flexibility in the response in some cases you may want it to be very flexible in other cases you should maybe like nope I want you strict answer every single time um call use case so uh this is actually a call transcript that you may want to look at so in this case we have a caller calling a uh an agent and you know they start out with hey how's it going you've been looking at my portfolio and all this so we can actually run this through here and let's go ahead and run it get all the answers back out and basically they're concerned about the performance of their portfolio so if you have someone where you need to get some uh you know themes or whatever out of a single call transcript you can actually in this case uh generate detailed call summary notes from the text that we put in in the call Prompt and then we get the call summary Chando called because the two discussed portfolio performance which has been underperforming the major indexes so that's great uh generate a list of topics discussed from the tall from the call and then basically it will generate a list of what the key elements were that were discussed in the conversation to save you quite a bit of time generate a list of follow-up tasks for the agent from the text so this could be something analyze the current portfolio develop Outlook so it actually has a lot of context about this call they're just making some pretty sound recommendations for what goes on and the last one how was the caller verified by the agent verification the agent verify the color we're asking caller's name in the last four digits of their account number uh this one's a fun one the disgruntled call case this one is an individual that got mad at the phone tree trying to get to an agent because they bought something that broke immediately after they plugged it into the wall not only did it break it started smoking um so they're calling the agent they're mad because they're soot all over their uh office now and they're getting ready for a triathlon so that's the gist of this angry caller that's calling in uh so what you can do instead you can read all that on your own but we'll go ahead and load this up uh so you can see it and uh we've got the transcript and the call summary is they're frustrated by the automated call tree and the amount of time it took to get through to a human being um this purchased a diver device like I said it started smoking when they plugged it in uh anyway so you kind of get the gist of what the summary of the call was it lists the mitigations the agent used in the call apologizing offering a replacement product expressing empathy offered to expedite delivery expressing appreciation uh wishing them uh good luck with their next race and then I you know thought it'd be fun it's like let's write a knowledge based article for future Service agents to see you know how open AI would respond it's like if they can figure out the context of how can I use this call to teach someone else that maybe how to mitigate an angry customer how to de-escalate and it did a really good job so this is something else you might want to think about using is the way to do that so but all right let's go and move on to our next topic okay so let's move on to our next topic so we're going to go back to our GitHub repo for useful repos and we're actually going to start with the third one chat GPT Enterprise data with Azure open Ai and cognitive search this is an architecture pattern that we're probably going to see more customers adopt or something similar to this so this is the loose architecture to where we actually have a user you know app we have an orchestrator which is going to be the application server now here's the thing when we think about open AI we probably are not going to have our users just we're not going to deploy an open eye endpoint and let the users run wild there is a scenario for that and we'll talk about it in a few minutes where you may just want openai to be available to your users to do document summaries of their own documents or um email summaries or help me generate an email things like that and we'll talk about an architecture pattern in a few minutes and how that would look so that if you still wanted you know something like open AI in your own Azure subscription where your data is protected that that may be necessary and we've got some actually really good interesting use cases for that scenario but more likely with several customers most of my customers that come to me with are like we have PDFs we have transcriptions we have things like that we want to create a system to where like for our HR documents for instance where our customers can actually or our users can actually come in and ask questions of documents that we have preloaded into something and that's where this pattern comes in because the way this will actually work is you would preload your documents take a PDF for instance run its reforms recognize or get the text out of it load the text index it in cognitive search so that you can run a query against cognitive search get the relevant documents back out pass that to open Ai and then see if we have an answer for the question or not so and that's kind of this pattern we're going to talk about the first pattern we're going to talk about so the data in this repo is in the like I said it's in the SQL ship so you can take a look at it but the data that we actually have is We have basically created a fake company we have the benefits options which is our Northwind Health Plus benefits our Northwind standard benefits perks plus which are the perks you get for working for the company basically it's like a thousand or fifteen hundred dollar reimbursement for gym equipment and gym memberships and you know Fitness devices things like that the employee handbook and the roll Library which is a list of jobs that are at the company so the way this particular repo works it will actually go through run through some scripts when you initially set it up run a series of Python scripts that will actually go in and take these documents divide them up into single Pages pass them to forms recognizer and then from forms recognizer they'll actually get uh indexed in Azure cognitive search so what will end up happening is I will end up with a blob storage account that is taken by multi hundred page documents and divided them into single pages and the reason for that is if we look at this this is what the repo looks like when you actually get it deployed and it's the GPT Enterprise sample and you've probably seen this before I think it's been done it's a Microsoft conferences but you know now you can actually go deploy this on your own if you so choose so you can ask it questions against the documents that we have loaded what's included in my Northwind Health Plus plan so we'll wait a second and it'll come back and but one of the things I want to be clear about is there are there's context here it didn't say what's included in my health plan what's included in my Northwind health plan that prompt engineering is an important thing to take note of because here's what's going to happen if we click on the light bulb when you do this demo and this is where you're going to need to start worrying about prompt engineering and how how users are interacting with the system it's searched for Northwind Health Plus plan benefits comparison standard okay which isn't exactly what I put in but basically it took this query that I put in the question that I put in and restructured it now you'll see there's a prompt here you are an assistant that helps company employees with their health care plans and questions about employee handbook be brief in your answers answer only with facts listed in the list of sources below if there isn't enough information below say you don't know do not generate answers that do not use the following sources this is called prompt engineering and this is something that you would actually do because if we go back because then the next question is where did that code come from so we're going to go back to the app we're going to look at the back end and basically we're going to look at these and you'll see that this is one thing you'll have to spend some time learning is Lang chain at some point because what will happen is if we look at the read retrieve for instance see if this is the one yep so we'll see that we've actually got some python library and this is actually the back end code that's running in an Azure app service right now that's doing the lookup for us so if we come back here and take a look at this we can see let's get that out of the way a little bit we can see here that we actually have the template prefix we which is what we saw just a second ago you are an intelligent assistant helping contoso Inc employees with their health care plan questions and then it goes through and you'll notice that each one of these approaches actually most of them have something to where it's trying to depending on which which package or which file we have to hit which function we have to hit you'll get a different prefix so it may be a little you'll notice some difference between the lane chain prefixes but what ends up happening is this will actually get passed in as part of the prompt which is what we saw right here that's where this is coming from so and then we'll also see that it came back with the document so it's like yeah here's the document that we found so if we can't pull that down there we go so citation this is the page where the information was actually stored so and they pulled this from blob storage because when the initial script ran that created this entire website it also went through and processed all the information we had about the data and that was where I told you that it read these documents in divided them into single Pages sent each page to form recognizer and then loaded that into the index search with a reference back to blob storage so that it would be able to do this and we can look at the supporting content because even though it just came up with the options two page there could be more content out there and then we can look at the thought process and figure out you know what were you thinking when you asked this question and ways to trip it up because I think that's important as well uh what roles are there I'm going to ask it kind of a vague question let's see if it can come back and it's kind of variable I'm sorry I cannot answer this question in the source is provided now I mentioned already that there is a roles Library out here that I should be able to look at roles and descriptions at contoso Electronics so we know that it was uploaded I uploaded it so let's look at the thought process employee health care plans is where it searched roles in the employee handbook so it was part way there so let's change that up a little bit to a different question let's just restructure the question what roles are there at contoso Electronics let's see if we can do a better job I can't answer that question provide this is where I've seen this uh and I've seen this repetitively let's see here provide a list of job roles at contoso I don't need a question mark let's see if we can get it here there we go so now we got it uh let's look at the thought process for this one see if it comes up uh rolls contoso Electronics employee handbook health care plan so until we actually narrowed it down see if that's the same text uh job roles contoso Electronics so the prompt engineering matters because we were trying to get for what it searched for and this is where you're going to have to do some testing and this is where you're going to need to do some testing with your users to find out what kind of questions are they asking uh one way I've kind of jokingly described uh chat gbt is a three-year-old with infinite knowledge and no context so you have to be very specific about how you ask it a question uh and you know more details and more context the better because you can see this is a good example that it kind of it kept missing the point until I finally broke it out just a little bit now once we've done that you can actually go in because since we have the role library and you can actually come in and say what does what does a uh and I didn't put do on there but let's see if it'll figure it out what does the vice president do the VP of research and development is responsible and we'll see that we've got the role library and the employee handbook so we can actually go through and get some context uh vice president of Human Resources that's a resource there we go research and development so that's the first page even though we didn't get the second page let's look at the other one the employee handbook number nine is where it found it uh vice president of products so it actually had two references for that so we'll come back and do that so this is a good way to get started uh in uh chat GPT just because uh this will kind of get you going to see how it behaves and search your own documents which I think is going to be really important many customers that have come to us and have asked it's like look we either have a document library or we have um transcripts so there may be a different architecture pattern depending on what you're doing and in the next part we're going to talk about the next few minutes we're going to talk about the way you can do this with other data and how the architecture pattern might look so the next architecture pattern we're going to take a look at is actually relatively easy we're going to do the cosmos DB and open AI chat GPT this one's relatively simple there's not us if we have an architecture here for it I don't think we do because basically it's just chat GPT with your data being stored in Cosmos which I kind of like you're like well and what's interesting about this one is we're not loading any documents in this might be a version uh let's imagine a version of the world where you gave your users access to chat GPT in a protected Azure subscription and you wanted them to have the abilities to go through and ask you know questions like you know for this case you know it was Harry Potter which we can all figure out we all know who Harry Potter is but there's no documents based on this but you know you may just kind of want to keep it as a basic search engine now here's where it could come in handy and email summarization or meeting summarization uh I had a I missed a meeting I'm not going to tell you what it was this past week it was about an hour and 10 minutes long uh I took the transcript from the meeting and I divided it up into 4 000 tokens so it's four sets of 4 000 tokens for a one hour ish meeting and I ran it through chat GPT to get the highlights of it much like the team's transcription I did my own version of that using this Cosmos and I'll show you why Cosmos is involved at all uh but that way I could actually summarize the meeting without me attending it I'm not encouraging to start skipping meetings and summarizing transcripts with chat GPT but you know if it happens it'll save you an hour of time from having to go back through the meeting and then if there's something interesting obviously you can go back but what I can do is remember uh from the Jupiter notebook a little while ago we had the Mad caller I took that transcript and we're going to pass it in uh summarize the following and we're going to pass that in and we can see that this came out with about 559 tokens and we can see the speaker is frustrated with the company's automated call system which is what we saw before so this looks very familiar uh so we still have this down here uh provide summarize the following in bullet points this might be useful for users to be able to uh come back and figure out you know what's a way to summarize large sums of text and let's just instead of doing that write an email choose the customer apologizing spelling if you misspell stuff in chat GPT shockingly it figures it out I can't tell you how grateful I am for that when I write an email to the customer apologizing for the issues and let's see if it can still keep the context and remember what we sent in um I'm writing to apologize for the issues you experienced whether automated call system and malfunction of your device so this is one where you might Empower your users to actually get a little bit more work done a little bit quickly I mean obviously writing that email is not going to take 10 minutes but it's like it's still going to take some minutes and it's going to take a context switch from whatever you're doing daily so this is kind of interesting now on the back end uh let's look at where this is because what I wanted to show you here is this is my Cosmos Cosmos DB account uh uh and you'll see that I've got basically the items that are being stored out here and you'll see that the name of the chat was Harry Potter because that was the first question I asked then I asked who is Harry Potter uh Harry Potter a fictional character is a fictional character so that was the answer it gave me and then I moved on to the following with the transcript now first off this particular deployment this Azure sample for Cosmos DB it's not multi-tenant which means that if you give this to two people they're going to be able to see each other's chats so you're going to have to do some coding to get this to be multi-tenant second part I like about it is for a b testing this gives you a good idea of what would my customers be asking and how are they asking questions that will help you figure out what kind of training you need for your environment or for your users and what kind of prompt engineering training because if you turn this over to a user they're still going to have to know a little bit about prompt engineering to make sure they get the type of answer or the type of response they want so that's going to be very important so I think there's a really interesting use case for the chat GPT backed by Cosmos to track your text and obviously Cosmos is just the database that ended up getting chosen you could write it to blob storage you could write it to redis Cache you could write it to you know wherever you wanted so I think that's really interesting all right let's take a look at our next one all right let's look at our last sample here so under our useful repos we actually have one called Azure open AI embeddings q a this is another uh basically another chat against your data which I really like but what's interesting about this one is they changed it up a little bit and this is kind of the way that uh I guess Azure open AI is seeing the world pre-build and post build if you want to see it that way or build conference I'm seeing a lot more information as of recently where we're using Vector databases to store the information about the documents that we have stored like before we had actually loaded the documents into a cognitive search index and had a reference to the document sitting in Blob storage uh they this one's done a little bit differently where you actually take your data set let's say we're using the same HR documents that we had before run them through forms recognizer potentially run them through Azure translator here which I like but what is different about this one is this product could be something else you could have instead of forms recognizer in Azure translator these could go from video to speech to text or to video indexer so that you have basically you know you could swap out any cognitive service there and that's what I like about this one so then you go to the Azure open AI service and get the embeddings okay so we actually do have an ADA embeddings model which will get the embeddings out of the document store those embeddings in a vector database now right now I think we've announced this if we haven't I'll probably get in trouble for saying this we do have Azure cognitive search Vector database coming out it's in private preview right now but I'm pretty sure we announced it at build but you could use pretty much any Vector database to do that including Cosmos postgres Vector which G8 last week which I'll say you know the week prior to June 17th because obviously my last week's not going to be here last week but and then basically what happens is when you get to the user end we actually have the users on this end ask a question it goes through the embeddings uh and then it hits the vector database and then goes to the answer prompt to get the answer you want back and respond back so I like this one a little bit different I like this one different than I like the others I think this one opens up some interesting possibilities for swapping out other cognitive Services instead of translator you could have video stored you could have audio stored in Blob storage and this could be the solution that you use so what this would actually look like is this is what you actually get deployed is the open Ai and you'll see and this looks like an administrator panel so I really like that too is that you could actually come in here take you this uh code make this your administrator panel so that you can actually write the code on the back end that goes against the vector database which in my case I'm using redis cache because if you drill into this if we go back we drill into this you'll see that there are several options for deploying this we'll go down with Azure cognitive search with Azure cache for redis uh redis stack I went ahead and did it with redis just because I thought that would be kind of interesting and the other thing I like about this one is this one click deploy so getting all of this data out there or getting all this infrastructure out there is really now the cosmos that we did in the last few minutes was also deployed to Azure so these are easy to get up and running so you can play with them and I really like that but what you get here is the open AI queries the chat uh in there you've also got add document which you can drag it drag and drop a document up to uh I think you can or just click browse uh up to you know uh 200 mag PDF jpeg or whatever add text to the knowledge base add stuff and batch I like that document viewer this is what I've already got loaded out here so we can actually see the documents I've already pre-loaded so I've got the benefits I've got the health benefits I've got the perks plus the employee handbook index management you can actually see how and where the data is stored a little bit so you can get some information about how this is set up there's a Sandbox where you can work on your own prompt engineering and see how that goes and there's a couple out here for basic summary bullet points explained to a second grader so you've got some other stuff you can do as well so but when you go to chat it's basically going to work like it did with everything else uh so you could you know what roles exist in contoso electronics since that's the popular question we asked before and it will actually hit the vector database instead of the cognitive search which in this case we get a short list but so that's not too bad and it's a good chat interface as well so this is more of a kind of an administrator type view of the world but you can take the code and do something else uh what roles exist provide so there we go see if we can't get a little bit more detail there uh yeah it's giving me kind of basic low end information it's like it's not giving me that nice bullet that we had before but it does give me some uh roll Library it does give me some links where I can actually go down and look at the actual text that was stored that way it's a little bit different the way it's set up but you can see that the last thing that it did was the job role so it did give me a list of what I was looking for even though it's not quite as elegant as it was in the other demo uh so yeah that one's the I like this one a lot uh I think it's interesting because it's a it's yet another way of all the things that I've shown you this is another way of uh basically doing documents uh on how they can actually get set up and work all right so the last thing we're going to talk about is the bottom of my repo here which is going to be blogs and posts uh as I find useful blogs and posts and there are tons of them out there but I try and pick ones that to try and minimize the uh the I don't want to send you a too long didn't read I could put 100 out here and it still wouldn't be all the ones that exist on the planet so I'm trying to put some interesting ones that are useful uh some of the stuff is from The Fast Track Team some of the stuff is going to be from other people within Microsoft this one I like the 10 generative 10 ways generative ai's transforming business where it actually talks about 10 different use cases and this might be useful for you to go through like content creation and design accelerated automation personalized marketing chat Bots and virtual assistants uh you know so this kind of gives you and the whole point of this video was to kind of give you the art of the possible we have yet to imagine every way a customer might actually use open AI in Azure uh so we're always excited to hear and learn about what other customers are doing as as always anyway I hope you enjoyed this video everyone have a great day

Channel: Introduction to Machine Learning and Data Science

Views: 5,428

Keywords:

Id: TI85JJVPnrM

Length: 37min 55sec (2275 seconds)

Published: Thu Jul 13 2023