Train your own Enterprise custom Data with Azure OpenAI Semantic Vector Search | Updated Feature

hey everyone welcome back in this quick video we are going to see the upgraded feature of azure open AI Studio this is a feature we were all waiting the vector search Yes you heard it right so earlier we were doing the full text search that is a simple search or either semantic search and now Azure open AI studio is back with a new feature where you can vectorize your data and then make a search on top of that and that too with your own custom data so let's get started so I'll go to our old feature that is add your data that is still in preview that means Microsoft is still upgrading and adding more and more features to this so the new feature that it has added is a vector search so here you can see and also you can give the indexer schedule that means if your data is keep on upgrading you can always run your indexer on schedule so I have selected Azure blob story that means I should have some data files over in my container I'll be choosing the storage account name the blob container okay since I do not have a Azure cognitive search resource I'll quickly create one let's pick up e2s and since most of the functionality is still not available in the free version we are going to go with the basic tile and select and review and create and there are other features also available in Azure cognitive search and Azure open AI service that is a private endpoint that I haven't included yet that comes mostly in the networking piece where no other resource or no other application can access your Azure open AI only certain networks or subnets will be able to access it so I'll need to learn them and yes I'll come back to you with a new video on that but this is the most important piece how to create vectors on your data so let me refresh it and here is our resource and let me give a new index name oh sample idx okay an indexer yes you can schedule it once hourly daily uh if you do it from the cognitive search itself they also provide a custom scheduler as well so here only three options are available but you can customize this from the cognitive search Resource as well so let's keep it at once only I just need to run it once and in order to use the vector search you must have a embedding modeled deployed on your Azure open AI that is currently it's not here let me just add that model that is text embedding header 002 and also if you don't find let's suppose you have access to gpt4 and of course that's the most accurate enough and also the costliest model if you don't have access to gpt4 that's fine if you have access to gpt4 and you don't see text embedding Eda under models that means the region that you have the Azure opener resource do not support both the services so let me tell you what I have meant by that there is a table here under the official documentation which tells us which region certain models are supported here you can see gpd4 models are available in this region and the 32 000 context is available in this regions uh in one of my client what I did I installed it in East us uh the gpt4 but I found that the embeddings model is not supported in East us here you can see oh yeah sorry East us2 it was East years 2 that I have worked on but my meetings model is not supported there East years too so I had to change it to Canada east or yeah I had to change this to I think Canada for it to work okay so if at all you don't find certain models that you would like to be included make sure you check out this page and see what models are supported in what regions okay um so something is going to retire version zero three zero one of GPT 3.5 Turbo okay yeah let's come back here and let's deploy this model I'm going to use the same name it's a good practice that you use the same name because if you have multiple deployments uh you'll not get confused that which is the model that I have deployed okay so I have it like gpt4 32 000 context and text embedding okay let's go back and restart the process okay everything is selected perfect now I can choose the embeddings model okay okay I need to change the tokens per minute limit of my embeddings model how you can do that go to your deployment go to your edit deployment Advanced option and increase the limit uh currently it asked me to make it fifty thousand but let me do it at the full limit that is supported under my subscription similarly you can do it for the other ones as well if you like to increase the tokens per minute limit then you can do it on all your models and also you can do it while deploying the model as well let's suppose I am deploying this and you can click on Advanced option and you can change the limit while deploying as well or if you have already deployed make it edit deployment okay perfect now now let's go back and try once mode okay I acknowledge and next search type so this is most important well each of the documents that you have the data you have the pattern in your document that you have varies everyone has their own set of document and it's Unique okay so now which strategy works best for them you need to try and check which is the best strategy for them first of course you can start with the simple search then you can go with the semantic search after that the third strategy is Vector search fourth strategies Vector plus simple and the final strategy is Vector plus semantic these are the five strategies that you can try and which strategy works best for you go fixed with it yes so let me show you that table that I am referring to and one more new feature has come power you can connect Azure open AI to power virtual agents that is your own data you can connect it to power virtual agents and this is a this was I was talking about private endpoint support but let me show you that table the five strategies yeah this is the one first always try with a simple there is no additional cost it's a keyword based search if you find it better go fixed it next is semantic search so here you can see benefit of that uh yeah of course the simple is the most fastest search so if you need a faster result you fix two simple search next is semantic uh it's kind of it semantic meaning of query terms so if you have some synonyms as well so it will be able to uh recognize that okay but semantic search is costly it has its own separate pricing tire that and the freight Rising Tire provides you thousand requests per month but if you go over that limit the pricing is 500 per month it is not upfront it will be daily chargeable so next strategy is Vector search it also has an additional pricing the additional pricing is for the embeddings model because each request when comes in it converts into the vectors using the embeddings Ada model next strategy is Vector plus simple it's a inbuilt feature that you can support or if you want to customize it there are sdks and client libraries available in both C sharp and python still in beta but they are available and final Vector plus semantic plus simple this is the most advanced one but it varies of which model is best for you you need to try it and then fix on that sometimes if you have a very straightforward document with no sir with very simple queries then just simple will work best for that if you have some users who just type some long sentences we should understand the meaning of that context then yes the vector search will come into picture here also it says the optimal search option can vary depending on your data set and use case you may need to experiment with multiple options to determine which works best for your use case so there is no best and least best model it's up to your data set and use case whichever box works best for you then pick that up okay I'm going to select the last Model just in case just want to show you how it works because simple and semantic we have already seen in our last videos Vector is very simple and basically all the three options we see it here so I am going to pick up the vector plus semantic before that I need to enable the semantic search I'll go here semantic and select plan okay now it's fine I can choose Vector plus semantic here also it says extra pricing will be for the semantic search and extra pricing for the vector embeddings that is embedding's Ada model so I acknowledge that next save and close let me see if I have any documents in that storage account I remember I added them but let me cross verify oh I don't have any documents let me quickly upload I'm going to pick up our old standard document that is quantum Computing it has two pages it is a very simple document well the simple search will work best for these type of documents but I am going to show you how vectors will be included here okay now let's go back let me remove it I'm not sure because I started the process and then added my document okay let me quickly check my index if my document is there or not yes it's not there Okay so now it's good good time for me to utilize one more feature that is the indexer I can rerun my indexer okay indexer not found strange I think that might be a reason because uh I have not added any document that's fine let me quickly add it again so there is another feature Microsoft Azure open AI can add is how many chunks we can create and what should be size of each of the chunk and also what sampling that is the olap we can add and how many results we can retrieve so all this customization can be performed maybe in the future Microsoft will add them because I have seen the scripts behind the scene which is running behind this scene uh there it is all mentioned uh the chunk size the overlap size so everything is mentioned there maybe if we can if we have that customization here in Azure open AI Studio then there is no need for us to code our custom requirement everything we can do it right from here maybe not now but maybe in the future yes Microsoft will definitely add them here you see um this is a thing even I am seeing a quite I think this is the second time I have seen while the process is running it is kind of creating two indexes uh one is this extra index actually Dash index I am sub I'm just assuming that it's a temporary index that Microsoft creates and what it does maybe from here to here it loads the final data but I'm not sure uh even when I was running one of my scripts it kind of created this two indexes one is a temporary one um but need to find out why is that but you automatically uh after indexing the temporary index will get deleted and we'll only see this one that we have created and let's see if any indexer is there yes and we have two indexers that's range we have two data sources that is also strange okay I guess uh it is storing chunks somewhere in another container I need to check my storage account nothing is here let me check my table nothing is here maybe something is happening behind the scene okay that's fine okay let's Now quickly check our index how many chunks it might have created let's refresh okay here is the data okay so in the earlier videos and earlier implementation we only had this many fields content file path title URL and the ID and the chunk ID now they have the content Vector as well there is one more field that you can manually add basically do it from the script that is a title vector that strategy also you can follow it's there in Azure cognitive search documentation uh in that examples they have included two Vector Fields one is a title Vector one is a Content vector here using the open AI Studio we have only content vector okay I'm assuming that it has created uh there is very less data in the second chunk but that's fine it's up to your data size and let's go to indexer okay ah since it's a one-time indexer so it is deleting them okay okay now I understood why it's not staying here if you have scheduled the indexer on an hourly or monthly basis then yes you will see here indexer and you can manually trigger them as well and since this is a manual process kind of a one-time index there is no data source associated with it okay since we are done with this let's ask our questions uh here in this data set there will not be much of a difference I'm just showing you a process how it works it's best if you try it on your data set because here it will not make much of a difference yeah and also I am using gpt4 yes it might understand my system from better if you if I write any I'm just keeping it as default system prompt and here you can see your citations okay this is just the process I am showing you the actual implementation may vary based on your data set and also the strategy Also may vary okay it's giving me long and long answers I can also limit that through my system prompt or I can just limit max response to some 100 tokens okay perfect uh this is how it works so behind the scene what is happening let me show you a diagram I think I will not have but let me just verbally tell you you ask query query is converted to vectors during you using your embeddings model then vector similarity search is done on your cognitive search index it gets some data the top results and then Azure open AI will just formulate a response based on your top result and show it to the user okay so this is the process that it is happening and this is a very good feature and it will have the best accuracy since Microsoft has introduced the vector search it will have the best accuracy in getting back the results okay thanks everyone see you hope you try this out and if something uh some issues some extra implementation then of course you can get in touch with us and uh also we have added a new service uh that is database migration microsql server migration we'll be doing you can check out our services page and and also our pricing on that yeah okay thanks everyone have a great day

Channel: Dewiride Technologies

Views: 1,604

Keywords: Azure, Open AI, Search Demo, ChatGPT, enterprise data, training, chatbot, AI, machine learning, NLP, natural language processing, artificial intelligence, tutorial, setup, custom knowledge, cloud computing, Azure services, Microsoft, azure search, openai, bot, train own data chatgpt, chatgpt train custom data, use own data in chatgpt, enterprise data in chatgpt, cognitive search, vector db, train custom data, chatgpt train own data, azure open ai train data, open ai

Id: 1cezdL3Ia-E

Length: 23min 13sec (1393 seconds)

Published: Sun Sep 03 2023