WEBVTT 00:02.760 --> 00:07.720 In this tutorial we are going to build our first rack agent. 00:09.200 --> 00:14.320 So rack stands for Retrieval Augmented generation, which basically means we give the large language 00:14.320 --> 00:17.440 model access to the external knowledge. 00:18.560 --> 00:23.880 So any kind of data you want usually that will be your own internal data or documents. 00:25.040 --> 00:32.600 So rack applications are kind of like custom gpts where you train ChatGPT to respond based on your on 00:32.640 --> 00:38.320 your own content or files instead of just its pretrained knowledge. 00:39.080 --> 00:46.720 And that's what's awesome about this setup, is that you can upload documents straight into Google Drive, 00:47.240 --> 00:51.000 and the chatbot will automatically learn from them. 00:51.160 --> 01:00.720 So every time you add something new, it will be vectorized to our vector database Is using this part 01:00.720 --> 01:05.320 of the workflow and instantly updates its knowledge base. 01:05.560 --> 01:10.560 This is incredibly useful for keeping your agent up to date without any manual work. 01:11.360 --> 01:13.600 So it will be done all automatically. 01:13.880 --> 01:15.720 We will use a practical example today. 01:15.720 --> 01:24.920 So let's say you are interested in the latest features introduced with iOS 18 from Apple, and you simply 01:24.920 --> 01:33.600 upload the official PDF document into your Google Drive, and the chatbot can immediately answer questions 01:33.880 --> 01:35.280 based on that content. 01:35.760 --> 01:40.360 So Apple could use this agent and embed it into their website. 01:40.360 --> 01:50.120 And any time a potential customer has questions about the new iOS 18 features, it could instantly access 01:50.120 --> 01:53.720 the vector database and provide an accurate answer. 01:53.720 --> 02:02.250 So that's why ROC agents are especially powerful for customer support as they give users real time answers 02:02.250 --> 02:05.970 based on the latest internal docs or updates. 02:06.530 --> 02:10.730 All right, so first open your workspace and create a brand new workflow. 02:11.730 --> 02:12.890 So create workflow. 02:13.610 --> 02:17.650 And let's start by adding our Google Drive to the node. 02:17.650 --> 02:21.890 So search for Google Drive. 02:26.610 --> 02:33.010 And we want to select the option that says on changes involving a specific folder because we want to 02:33.050 --> 02:34.010 automate the process. 02:34.010 --> 02:39.130 So when we add a new document into this folder this workflow will be triggered. 02:39.450 --> 02:45.730 So this node will watch drive folder and triggers whenever new files appear. 02:47.730 --> 02:51.010 Here of course we need to connect our Google Drive account. 02:51.050 --> 02:59.730 So if your Google Drive is not connected yet I click on this drop down create new credential And here 02:59.730 --> 03:02.250 you have to provide our client ID and client secret. 03:02.450 --> 03:08.810 If you don't know how to find those, go to the How to set up credentials section and you will find 03:08.810 --> 03:14.210 a separate tutorial on how to connect Google Apps into your account. 03:14.810 --> 03:21.570 If you if you have connected to your Google apps before, you can just simply sign in with Google to 03:21.610 --> 03:23.610 refresh the connection just in case. 03:24.210 --> 03:25.290 Now we can close it. 03:25.770 --> 03:30.730 Now we want this trigger to watch for our Google Drive folder every minute. 03:31.530 --> 03:35.010 But you can change it to every hour, every day. 03:35.650 --> 03:37.090 It's completely up to you. 03:37.450 --> 03:45.050 But in that case, I want this trigger to check my Google Drive folder every minute so it will be easier 03:45.050 --> 03:50.970 to test our automation and agent in the future steps. 03:51.250 --> 03:58.330 Now we want this automation to be triggered based on the changes involving a specific folder, so you 03:58.330 --> 04:01.930 can leave it as it is, but you can also change it to. 04:03.290 --> 04:09.090 Activate the workflow when it will notice changes to a specific file. 04:09.730 --> 04:11.050 Within your Google Drive. 04:11.530 --> 04:18.930 Now we want to create a dedicated folder in Google Drive to store this document. 04:19.210 --> 04:26.090 So I'm going to attach the iOS 18 new features document as a resource. 04:26.530 --> 04:28.130 So go to your Google Drive. 04:29.490 --> 04:30.570 Create a folder. 04:31.210 --> 04:36.450 I call mine First Rack Agent and then go to the new. 04:36.530 --> 04:41.730 Click this plus button and click on File Upload. 04:42.010 --> 04:46.170 And you can upload the same document or whatever document you want. 04:46.570 --> 04:52.450 Now when I click on this from list folder so we can select the folder and it should pop up. 04:52.850 --> 04:54.210 I can just search for it. 04:58.700 --> 05:00.980 As you can see, it appears here. 05:01.460 --> 05:02.740 So first rock agent. 05:03.980 --> 05:05.620 And we can watch for. 05:06.220 --> 05:09.460 And here we want to watch for file created. 05:09.460 --> 05:13.500 So when a file is created in the watch folder select that. 05:13.940 --> 05:17.820 And now we want to fetch the the test event. 05:18.060 --> 05:23.660 Because basically what we want to do we want to fetch the document we just uploaded. 05:24.300 --> 05:27.140 And we can see if the data come through. 05:29.860 --> 05:30.900 Yeah as you can see. 05:33.020 --> 05:34.700 As the output you have our data. 05:34.700 --> 05:36.340 We have our document right here. 05:37.300 --> 05:40.820 You can switch to the JSON format or to the schema. 05:44.780 --> 05:46.620 And everything should work perfectly. 05:49.180 --> 05:56.820 So yeah, it's a good practice to ensure that N810 pulls data about your your PDF file. 05:57.180 --> 06:00.180 So you should see the file name file file ID. 06:00.780 --> 06:03.940 So let's look for the file name. 06:13.980 --> 06:14.820 That is here. 06:15.260 --> 06:17.940 And also the filed is right here. 06:18.180 --> 06:19.620 And fetch a lot of more data. 06:19.620 --> 06:21.700 But we are interested just in a few fields. 06:22.740 --> 06:27.620 So we know that the document was fetched successfully so we can move on. 06:28.420 --> 06:29.420 So let's close it. 06:29.420 --> 06:31.020 Let's add another node. 06:31.500 --> 06:37.940 Now every time we add a new document into our Google Drive folder, it will be fetched by this trigger 06:37.940 --> 06:38.380 node. 06:38.540 --> 06:41.300 Now what we need to do we need to download this file. 06:41.460 --> 06:43.580 So we need to add another Google Drive node. 06:46.180 --> 06:47.500 So search for Google Drive. 06:49.100 --> 06:52.580 And this time we want as an action we want to download the file right. 06:53.180 --> 06:54.140 So click on that. 06:54.500 --> 06:58.940 So we want to download a specific file and the operation. 06:58.940 --> 07:00.980 We want to leave it as it is. 07:00.980 --> 07:01.820 So download. 07:01.980 --> 07:05.980 You can also copy a file, create from text, delete, etc.. 07:07.740 --> 07:14.500 As you can see, after connecting this Google Drive node into our trigger node, the fetch data comes 07:14.500 --> 07:14.860 through. 07:15.180 --> 07:16.860 So we have it as an input now. 07:17.300 --> 07:23.220 Now this is very important because we don't want to select a specific file from from a list. 07:23.220 --> 07:26.500 So from our Google Drive we want to do it dynamically. 07:26.500 --> 07:33.100 So every time when we add a new file into our Google Drive folder it will be a different file. 07:33.140 --> 07:33.700 Right. 07:33.700 --> 07:37.380 So we need to add the file ID. 07:38.380 --> 07:42.380 So let's search for filed and we should find it. 07:42.700 --> 07:43.740 Add the spices. 07:44.260 --> 07:49.340 So let's search for this spices object. 07:54.180 --> 07:57.390 And as you can see here under spaces, we have our field. 07:57.830 --> 08:00.670 And the name of this file is exactly what we are looking for. 08:00.710 --> 08:05.110 So our iOS 18 all new features September 2024. 08:05.150 --> 08:08.950 So let's simply drag this ID and drop it right here. 08:10.030 --> 08:15.150 So now as you can see we successfully mapped the file into this field. 08:15.150 --> 08:16.430 So that's it for this note. 08:16.470 --> 08:17.390 Now let's move on. 08:17.910 --> 08:18.150 Yes. 08:18.150 --> 08:28.870 So we simply used the dynamic ID coming from uh this note node output to the ID field to download the 08:28.870 --> 08:31.150 correct file using this node. 08:31.310 --> 08:33.950 So now let's uh just rename this node to. 08:38.190 --> 08:38.990 Root file. 08:42.270 --> 08:47.590 And also let's rename this workflow to to vectorize data. 08:52.390 --> 09:00.630 From Google or from Drive the Pinecone, because we are going to use pinecone as our vector database. 09:04.350 --> 09:11.310 And we are going to vectorize files we store in our Google Drive folder to the Python vector store. 09:11.550 --> 09:16.710 Now the next step is to load this file into our vector database. 09:17.430 --> 09:22.110 So I really like pinecone vector database for rack agents. 09:22.390 --> 09:28.750 So before we add a new node go to Google and search for Python. 09:41.710 --> 09:45.910 Login or sign up if you don't have an account yet I'm going to log in. 09:49.430 --> 09:52.190 And as you can see, pinecone is a long term memory for AI. 09:52.430 --> 09:53.990 You can start building for free. 09:54.150 --> 10:02.550 So basically you can build at least a few rack agents for for free without paying the monthly subscription, 10:02.550 --> 10:04.470 which is really great for testing purposes. 10:04.870 --> 10:08.670 Now, once you log in, what you need to do, we need to create a new index. 10:08.670 --> 10:15.350 As you can see, I already have a couple of indexes, but for you it will be it will be completely empty. 10:15.830 --> 10:25.390 So I'm going to create new index, give it a name and a ten like agent. 10:26.950 --> 10:31.030 And the next thing is to configure our index. 10:31.030 --> 10:36.110 So what we need to do we need to choose text embedding. 10:36.150 --> 10:37.110 Very small model. 10:37.390 --> 10:40.270 This is most cost effective OpenAI embedding model. 10:40.310 --> 10:45.270 This is really important because we are going to use the same embedding the same embedding model in 10:45.270 --> 10:47.430 our Na ten application. 10:47.950 --> 10:52.110 So click on that and create index. 10:54.200 --> 11:00.840 Now the index has been successfully created and you don't need to worry about those details, but it's 11:00.840 --> 11:10.880 useful to know that the dimensions means that each text is stored as a list of 1535 numbers that capture 11:10.880 --> 11:14.480 its meaningful for smart searching for our agent. 11:15.160 --> 11:19.160 Now, the next thing we need to do, we need to create our API keys. 11:20.040 --> 11:22.240 So click on this API key section. 11:22.720 --> 11:24.080 Create new API key. 11:25.160 --> 11:26.000 Give it a name. 11:27.200 --> 11:32.080 My case is an agent create key. 11:33.760 --> 11:37.160 Simply copy it and let's go back to N810. 11:37.560 --> 11:42.240 Now what we are going to do we are going to search for pinecone. 11:45.480 --> 11:47.480 So we want to add a pinecone vector store. 11:52.600 --> 11:52.800 It. 11:52.840 --> 11:55.840 Of course, as the action, we want to add a document to a vector store. 11:56.520 --> 11:57.480 So click on that. 11:59.240 --> 12:01.560 Now we need to connect to our pinecone account. 12:01.560 --> 12:03.120 So let's create a new credential. 12:04.880 --> 12:05.920 Create new credential. 12:05.960 --> 12:12.000 And let's face the API key we just created and hit save. 12:12.600 --> 12:14.880 As you can see credential successfully created. 12:15.520 --> 12:16.720 So let's close it for now. 12:17.400 --> 12:24.800 Now feel free to use other vector databases so you can use different vector stores for example Supabase. 12:25.320 --> 12:28.560 So let's just quickly connect these nodes. 12:29.960 --> 12:35.600 And if you want to use other providers you can search for vector. 12:39.720 --> 12:41.400 And here we have simple vector store. 12:41.440 --> 12:43.240 We have a quadrant we have supabase. 12:43.760 --> 12:47.400 But for RAC agents I really like to use pinecone. 12:47.400 --> 12:49.360 So I'm going to stick with that. 12:50.840 --> 12:52.680 Now of course we need to configure this node. 12:52.680 --> 12:54.400 So let's open it. 12:55.760 --> 12:59.040 And the operation mode is of course insert documents. 13:00.800 --> 13:03.560 So we are going to leave it as it is. 13:04.280 --> 13:05.560 And now you have to. 13:06.160 --> 13:09.440 And now we have to select the pinecone index we have just created. 13:09.760 --> 13:13.560 So from the list I have created an agent. 13:13.560 --> 13:15.800 So I'm going to choose this one. 13:16.080 --> 13:18.640 You can leave embedding batch size as it is. 13:19.160 --> 13:21.120 And now we have to add the namespace. 13:21.680 --> 13:29.760 You can add the namespace directly in your Python account under databases right here. 13:31.840 --> 13:35.400 But I'm going to add the namespace in N810. 13:36.400 --> 13:40.280 So in this node I'm going to add option and Python namespace. 13:41.320 --> 13:44.680 And I'm going to call it rack agent. 13:47.200 --> 13:47.600 Right. 13:47.640 --> 13:48.200 That's great. 13:48.240 --> 13:51.370 Now let's close it safe. 13:51.850 --> 13:53.210 Now let's test our workflow. 13:57.650 --> 14:03.690 As you can see, we got an error because we have to add Embeddingmodel and we have to add document loader. 14:04.170 --> 14:05.130 So let's open it. 14:06.730 --> 14:10.170 But we successfully downloaded the file. 14:10.530 --> 14:12.170 We add it to our Google Drive folder. 14:13.130 --> 14:16.850 So now let's close it down and let's add the Embeddingmodel. 14:17.090 --> 14:19.010 I really like OpenAI's embeddings models. 14:19.010 --> 14:20.610 So I'm going to choose this one. 14:22.770 --> 14:29.290 And here you have to ensure that the model you are selecting is the same model we use in the Python. 14:29.290 --> 14:32.850 So it was text embedding free small model. 14:33.530 --> 14:39.130 You have also access to the large model and Ada 002. 14:40.490 --> 14:43.930 Also make sure that your OpenAI account is connected. 14:44.530 --> 14:45.770 And that's it for this node. 14:45.770 --> 14:48.370 So let's close it down and move on. 14:50.650 --> 14:51.850 Now we have to add document. 14:51.890 --> 14:56.690 As you can see, we are going to use the default data loader. 14:56.690 --> 15:01.010 So we want to load data from previous step in the workflow. 15:02.610 --> 15:04.690 So in that case from the download file. 15:05.970 --> 15:07.170 So let's choose this one. 15:07.450 --> 15:14.490 And as you can see as the input we fetch the iOS 18 file. 15:14.810 --> 15:17.810 And this is in the PDF format. 15:18.290 --> 15:23.490 So you can view it just by clicking on this button. 15:27.490 --> 15:31.450 This in JSON format the table schema. 15:31.890 --> 15:35.210 So we downloaded the file in a PDF format. 15:35.250 --> 15:37.050 So this is not the JSON format. 15:37.290 --> 15:39.450 So we need to change the type of data. 15:39.930 --> 15:41.250 It's going to be binary. 15:41.250 --> 15:44.530 So we want to process the binary data from the previous node. 15:45.050 --> 15:46.090 So let's click on that. 15:46.650 --> 15:46.970 Yes. 15:46.970 --> 15:51.810 So we are using the PDF, uh, document in this example. 15:52.130 --> 15:55.290 We can leave mode as load or input data. 15:55.330 --> 15:59.410 You can also switch over to load specific data if you want to. 15:59.850 --> 16:05.330 So this is this will be load data from a specific field in the parrot agent or chain which is a little 16:05.370 --> 16:06.450 bit more complicated. 16:06.930 --> 16:08.210 So let's stick with that. 16:08.210 --> 16:12.090 And as a data format you can leave it as it is. 16:12.130 --> 16:14.730 So automatically detect by Mime type. 16:15.090 --> 16:22.250 But you can also switch to the PDF because we are using the PDF document. 16:22.650 --> 16:24.090 So load PDF document. 16:24.130 --> 16:24.330 Yes. 16:24.330 --> 16:32.770 So if you want to sometimes upload different type of documents such as docs or uh, CSV files into your 16:32.810 --> 16:35.930 Google Drive folder, you can use this option. 16:36.250 --> 16:38.250 But in that case we are going to use PDF. 16:38.770 --> 16:39.890 Now let's close it down. 16:40.210 --> 16:42.450 Now next we need to add the text splitter. 16:42.610 --> 16:48.740 So we are going to use the accuracy for character text beta which means that it will split text into 16:48.740 --> 16:50.980 chunks by characters recursively. 16:51.420 --> 16:53.860 And it's recommended for most use cases. 16:54.140 --> 16:59.020 I recommend using it for RAC agents for better accuracy. 16:59.140 --> 17:00.700 So we are going to use this one. 17:01.380 --> 17:10.860 Now what this recursive character text splitter does is it breaks long text into smaller chunks so they 17:10.860 --> 17:15.100 can be processed or embedded properly by LMS. 17:15.500 --> 17:24.380 So no overlap means each chunk is separate and does not share text with the next one. 17:24.420 --> 17:31.780 So for this size of document, around 900 characters um with an overlap of 50 works really well. 17:31.980 --> 17:38.820 So it keeps each chunk short enough for processing, while the overlap helps preserve context between 17:38.820 --> 17:43.060 chunks, which is great for better search and retrieval accuracy. 17:43.100 --> 17:50.660 So here we are going to enter 100 and check overlap 50. 17:51.460 --> 17:58.100 Now let's test it and see if we are able to upload this file into our vector store. 17:58.540 --> 17:59.660 So let's close it. 17:59.700 --> 18:00.700 Let's hit save. 18:00.700 --> 18:04.420 And let's go quickly back to our icon. 18:05.740 --> 18:10.260 When you click on this index, as you can see there is no records yet. 18:11.380 --> 18:17.660 And the record is an object to add to an index which contains vector and some metadata. 18:18.180 --> 18:27.180 So now when we go to our workflow and click on Test Workflow, it should populate our namespace with 18:27.180 --> 18:27.700 the vectors. 18:27.700 --> 18:32.860 So it should split the document we uploaded into our Google Drive folder. 18:33.140 --> 18:39.980 As you can see, it keeps executing because it's waiting for us to create an event in Google Drive. 18:39.980 --> 18:44.460 So to upload the new file into our Google Drive folder. 18:44.660 --> 18:47.660 So what I'm going to do I'm going to pin the data. 18:49.140 --> 18:50.140 So let's stop it. 18:51.940 --> 18:56.740 I don't need to upload a new file into our Google Drive. 19:03.980 --> 19:05.300 So I'm going to pin it. 19:14.540 --> 19:24.940 Oh, it actually successfully vectorize our file and added it to our Python vector store. 19:25.860 --> 19:28.900 As you can see we have a ten hits now. 19:30.460 --> 19:35.900 So as you can see our document has been splitted and embedded into our vector store. 19:37.820 --> 19:40.060 We have 80 documents. 19:41.060 --> 19:48.830 And just to confirm as you can see we have 80 items as an output from our Python vectors to node. 19:49.190 --> 19:54.630 So we successfully downloaded the file from our Google Drive folder. 19:54.790 --> 20:01.670 So what we have done here is the automation was triggered when a new file was added to our Google Drive. 20:01.950 --> 20:08.510 It downloaded the file, then passed it to the default data loader to extract the content. 20:08.790 --> 20:14.790 That content was split into smaller chunks using recursive character text beta. 20:15.590 --> 20:19.790 Those chunks were then sent to OpenAI to generate embeddings. 20:20.350 --> 20:27.190 And finally, all the embeddings were stored into in our Python vector store for semantic search. 20:27.550 --> 20:35.110 So our agent will be able to go through those embeddings and do semantic search technique to retrieve 20:35.110 --> 20:39.110 the most relevant information based on the questions from the user. 20:39.470 --> 20:46.150 Naturally, you will want to upload more than one file into your Google Drive to improve the knowledge 20:46.150 --> 20:48.190 base for your RAC agents. 20:49.190 --> 20:54.510 So if you are planning to add more data inside your Python vector store. 20:55.230 --> 20:58.950 So Python vector database for example. 20:59.310 --> 21:03.950 So let's say Apple just released a new version of this document. 21:05.910 --> 21:11.670 And he would and he would probably want to drop that into your uh drive folder. 21:12.270 --> 21:17.510 So there is a way to keep everything more organized and easy to track. 21:17.750 --> 21:21.550 So you can attach metadata in any workflow. 21:21.830 --> 21:28.950 So that way, once the document is stored in your vector database, you will see the actual file name 21:29.150 --> 21:29.870 right here. 21:31.070 --> 21:33.590 So this is super useful for staying organized. 21:33.590 --> 21:36.390 So I will show you how to set that up in a second. 21:36.710 --> 21:43.390 But also quick tip if you are using your RAC agent to handle completely different topics and unrelated 21:43.480 --> 21:44.200 documents. 21:44.480 --> 21:48.360 It's a really good idea to create separate namespaces for each category. 21:48.360 --> 21:50.640 So for each, for each project. 21:50.760 --> 21:56.320 So now if you want to add um metadata go to default data loader node. 21:58.720 --> 21:59.680 Under options. 21:59.720 --> 22:00.560 Add option. 22:01.680 --> 22:02.680 Add metadata. 22:04.440 --> 22:05.400 Add property. 22:06.200 --> 22:07.160 Give it a name. 22:07.200 --> 22:09.040 So I'm going to call it Apple. 22:10.680 --> 22:18.480 And you have to drop here the name of the file to fill out this dynamically. 22:19.400 --> 22:21.840 So switch to switch over to mapping. 22:22.200 --> 22:23.280 Then go to schema. 22:25.600 --> 22:26.760 Look for spaces. 22:32.880 --> 22:34.320 And right here we have a name. 22:35.200 --> 22:38.040 Let's let's drag it and drop it right here. 22:40.960 --> 22:42.800 So as you can see it's all green and results. 22:42.800 --> 22:43.480 We have a name. 22:44.400 --> 22:45.280 Let's close it. 22:46.520 --> 22:47.920 Just unpin this data. 22:50.880 --> 22:51.600 Let's save it. 22:51.640 --> 22:53.880 Now go to pinecone. 22:54.360 --> 22:57.040 I'm going to delete this, uh, namespace. 23:00.680 --> 23:04.120 So delete that agent. 23:06.280 --> 23:07.600 And delete namespace. 23:08.160 --> 23:11.040 And I'm going to run this workflow once again. 23:11.920 --> 23:15.360 So now I'm going to upload the same file to. 23:15.400 --> 23:16.320 My Google drive. 23:20.200 --> 23:21.040 So new. 23:23.600 --> 23:24.320 Upload. 23:35.080 --> 23:37.920 Now let's go back to our workflow. 23:42.760 --> 23:45.120 So you need to make sure that your workflow is active. 23:45.120 --> 23:50.240 So that way our that way our trigger will watch our folder. 23:50.920 --> 23:52.560 So let's make this workflow active. 23:54.640 --> 23:57.760 Confirm I uploaded the file twice. 23:59.600 --> 24:02.120 As you can see two uploads complete. 24:04.920 --> 24:06.480 Now we will go to executions. 24:08.720 --> 24:11.960 You can see that it successfully uploaded the file. 24:12.400 --> 24:14.560 It's all green so there is no any errors. 24:17.080 --> 24:19.480 Now let's go to our pinecone vector store. 24:21.480 --> 24:24.680 As you can see, we have a record of 160. 24:25.640 --> 24:26.760 So let's open it. 24:28.320 --> 24:30.920 Go to the browser, refresh it. 24:31.320 --> 24:33.920 And now you can see that the metadata has been added. 24:33.920 --> 24:40.520 So I can see the exact name of the file I just vectorize to my vector store in Python. 24:40.840 --> 24:41.170 All right. 24:41.170 --> 24:41.650 Great. 24:42.130 --> 24:48.130 So yes, if you want this automation to run every time you drop a new file into your Google Drive folder, 24:48.130 --> 24:49.850 make sure to activate your workflow. 24:51.210 --> 24:57.650 And once it's active, it will automatically pick up new files, process them, and store them in your 24:57.690 --> 24:58.970 icon vector database. 24:59.810 --> 25:03.290 We built our RAC agents knowledge base using pinecone. 25:03.930 --> 25:11.570 Now every time we add a new PDF file like the latest iOS 18 features to our Google Drive, the knowledge 25:11.610 --> 25:13.410 base updates automatically. 25:13.930 --> 25:18.770 So now we are going to build an actual chatbot that uses this knowledge to answer questions. 25:19.090 --> 25:24.970 So by the end, you will be able to ask your chatbot anything from your uploaded document and get clear, 25:25.490 --> 25:26.570 accurate answers. 25:27.210 --> 25:32.650 Now, before we dive into building the chatbot, there is something very important to keep in mind about 25:32.650 --> 25:35.370 how vector databases like pinecone work. 25:35.530 --> 25:41.090 So even though they are super powerful for storing and searching through document data. 25:41.970 --> 25:43.090 They are not unlimited. 25:43.090 --> 25:50.250 So over time, as you keep uploading new documents, your database can start to get cluttered or even 25:50.290 --> 25:52.410 hit storage limits depending on your plan. 25:52.690 --> 25:55.930 So it's a good idea to keep things clean and well organized. 25:56.370 --> 26:04.450 So don't just dump in every single file you ever created, especially if they are huge PDFs with images 26:04.610 --> 26:12.010 or unstructured content, because that kind of data is not ideal for llms to to work with anyway, and 26:12.010 --> 26:14.650 it can make responses less accurate. 26:15.050 --> 26:22.130 So instead, try to keep your documents, uh, relevant, uh, concise and well formatted. 26:22.410 --> 26:27.650 So if you are no longer using certain data, it's totally fine and recommended to remove it from the 26:27.650 --> 26:30.490 vector store to make room for new updates. 26:31.090 --> 26:31.570 So. 26:32.290 --> 26:35.570 So now let's move on and start building the actual agent. 26:36.170 --> 26:37.530 So creating workflow. 26:38.050 --> 26:40.250 Then we need to add our first note, let's show. 26:40.250 --> 26:45.250 Let's add the agent to search for an agent. 26:51.210 --> 26:52.290 Now let's close it. 26:52.890 --> 26:54.490 Let's add our chat model. 26:55.330 --> 27:00.930 So we're going to use OpenAI model GPT four. 27:00.970 --> 27:03.930 I recommend to use GPT four for more accuracy. 27:20.010 --> 27:20.530 All right. 27:20.530 --> 27:23.170 So make sure you OpenAI account is connected. 27:23.170 --> 27:29.450 But it should be automatically connected if you set up your credentials in the workspace and in the 27:29.450 --> 27:29.890 project. 27:29.890 --> 27:31.810 So let's close it for now. 27:32.930 --> 27:39.180 Now we need to give our agent clear instructions on how to respond. 27:40.060 --> 27:42.100 So let's configure this node. 27:43.180 --> 27:45.180 Now let's add system message. 27:46.140 --> 27:47.980 So here we have to add our prompt. 27:48.780 --> 27:52.860 Let's switch over to expression mode and expand the editor. 27:55.340 --> 27:58.020 I'm going to paste the prompt I have prepared earlier. 27:58.380 --> 28:06.500 You can simply create the prompt with the help of ChatGPT or cloud, or you can modify this one. 28:06.500 --> 28:07.980 This is just a template for you. 28:09.300 --> 28:12.060 So let's say we can improve it. 28:13.580 --> 28:15.220 So you can say this is the role. 28:15.580 --> 28:21.500 So you are an assistant trained to answer questions specifically based on the iOS 18 feature documents 28:21.500 --> 28:23.900 stored in the Connected Vector database. 28:24.460 --> 28:25.620 And as the objective. 28:29.660 --> 28:34.860 Your job is to help users clearly understand the new features, settings, and changes introduced in 28:34.860 --> 28:41.060 iOS 18 by retrieving and referencing only the information available in the documents stored in the vector 28:41.060 --> 28:41.420 store. 28:41.780 --> 28:47.100 The user asks something that is not included in the uploaded content, or if the answer is not clear 28:47.100 --> 28:53.820 from the available data, let them know that the information is not currently available and also we 28:53.820 --> 28:55.820 can call it rules. 28:57.980 --> 28:59.460 So be clear and direct. 29:00.540 --> 29:10.860 Avoid Overexplaining I would say also use 3 to 5 sentences. 29:13.380 --> 29:17.740 Now with bullet points or short paragraphs when listing multiple features. 29:18.540 --> 29:23.020 If a feature only applies to certain devices or release phases like beta stable, make them clear. 29:23.020 --> 29:26.780 When mentioned in the comments, you can just remove it. 29:28.140 --> 29:29.300 Never guess or assume. 29:29.300 --> 29:33.020 Stick strictly to what's inside the vector database content. 29:33.020 --> 29:40.460 This is really important and my objective If is deliver helpful, accurate and easy to understand answers 29:40.460 --> 29:51.140 about the iOS 18 using only the documents, let's say answers using only the documents available in 29:51.180 --> 29:51.900 the vector store. 29:51.900 --> 29:55.060 So no speculation, no extra assumptions, and no external sources. 29:55.460 --> 29:55.780 All right. 29:55.780 --> 29:56.380 Brilliant. 29:57.540 --> 30:00.340 This prompt should work perfectly fine. 30:01.620 --> 30:02.700 Now let's close it. 30:04.660 --> 30:09.180 Now, to make our agent interactions feel natural, let's add a memory. 30:10.220 --> 30:11.660 So let's add simple memory. 30:13.820 --> 30:17.220 And let's set it to the last ten messages. 30:21.980 --> 30:24.980 So now your agent will really remember context. 30:25.500 --> 30:25.860 All right. 30:25.860 --> 30:26.380 Great. 30:27.140 --> 30:34.780 So for example, without memory, the agent will forget details immediately. 30:34.780 --> 30:41.670 So When I say hi, my name is Damien and then ask, what's my name? 30:41.870 --> 30:49.550 The chatbot will not be able to answer this question, but once he added the simple memory, it will 30:49.550 --> 30:52.590 be able to remember my name. 30:53.150 --> 30:56.270 So now when I test it and I open the chat. 30:57.750 --> 31:02.910 Hi, my name is Damien, I like dogs. 31:08.350 --> 31:09.670 Hi, Damien, it's nice to meet you. 31:09.670 --> 31:16.630 If you have any questions about the new features or changes in iOS 18, feel free to ask what's. 31:21.310 --> 31:21.910 My name? 31:27.230 --> 31:30.550 What animals I like. 31:33.630 --> 31:36.270 You mentioned that your name is Damien and that you like dogs. 31:37.070 --> 31:37.350 All right. 31:37.350 --> 31:37.870 So great. 31:37.910 --> 31:39.670 This is how the memory works. 31:39.670 --> 31:41.990 So this is great. 31:41.990 --> 31:48.790 And and because of the memory, you can sometimes feel that you chat with the real human. 31:49.430 --> 31:49.950 Okay. 31:50.270 --> 31:51.430 So let's close it for now. 31:52.590 --> 31:58.030 Next thing we need to connect the vector database. 31:58.710 --> 32:02.950 So our agent can pull information directly from our documents. 32:04.070 --> 32:13.270 We just vectorized and load it from our Google Drive folder into our pinecone vector store. 32:13.710 --> 32:23.390 Actually, let's check if I ask our agent to about anything included in the iOS 18. 32:25.470 --> 32:26.710 Features document. 32:28.470 --> 32:34.590 It will be hallucinate or just state that is not able to answer these questions because it doesn't have 32:34.670 --> 32:41.830 this information, access to this information as we want because we just added it to our prompt. 32:44.190 --> 32:44.870 Right here. 32:45.470 --> 32:50.030 So if a user asks something that is not included in the uploaded content, etc.. 32:50.870 --> 32:54.150 So now let's test our agent again. 32:55.030 --> 32:56.550 Let's ask it about. 33:00.910 --> 33:02.630 What's in the playground. 33:04.270 --> 33:05.270 Let's copy it. 33:15.150 --> 33:21.030 As you can see, the information about the playground is in iOS 18 is not available in the documents 33:21.030 --> 33:21.510 I have. 33:21.550 --> 33:29.670 And this is what we expect because we have not connected our pinecone vector database to our agent. 33:29.670 --> 33:33.150 So now we are going to now we are going to add a tool. 33:33.670 --> 33:39.440 We're going to add the search for vector. 33:41.320 --> 33:43.120 And this is the tool we are going to use. 33:43.120 --> 33:45.920 So vector store question answer tool. 33:45.920 --> 33:48.600 So it answers questions with a vector store. 33:49.880 --> 33:53.800 So simply add it and let's configure it later. 33:56.840 --> 33:59.200 Now what we need to do we need to add the vector store. 33:59.200 --> 34:01.520 We just we have created in the previous steps. 34:02.720 --> 34:05.760 So we are going to add our vector store. 34:08.280 --> 34:14.320 If you have created your own vector store earlier it will be connected automatically as operation mode. 34:14.320 --> 34:17.720 We want to retrieve documents as a vector store for training tool. 34:18.440 --> 34:20.160 So we are going to leave it as it is. 34:20.640 --> 34:24.440 And here we are going to select our Python index. 34:27.360 --> 34:30.000 So it was an RAC agent in my case. 34:32.040 --> 34:32.920 Okay great. 34:33.760 --> 34:37.640 So now as you can see you also need to add embedding model. 34:38.160 --> 34:41.760 So we are going to use OpenAI embedding model. 34:42.240 --> 34:49.120 The same you used for our Python vector store which is text embedding free small model. 34:50.760 --> 34:51.560 Close it. 34:53.120 --> 34:55.480 Now you have to add our model. 34:58.960 --> 35:01.520 So I'm going to use GPT for all. 35:04.240 --> 35:06.000 I'm going to switch to GPT for all. 35:08.640 --> 35:09.480 Okay great. 35:10.320 --> 35:11.760 And now we can configure this node. 35:13.800 --> 35:16.080 So now we have to give it a name. 35:16.080 --> 35:18.800 So let's call it iOS. 35:22.000 --> 35:22.720 18. 35:26.080 --> 35:32.960 And also the name of this node will be the same to name it. 35:33.480 --> 35:38.280 And here we have to specify when we want to call this tool by our agent. 35:38.960 --> 35:40.800 Now we have to add the description of the data. 35:40.800 --> 35:46.600 So when you hover over for this question mark you can see that we need to describe the data in the vector 35:46.600 --> 35:46.920 store. 35:46.920 --> 35:50.040 So this will be used to fill this tool description. 35:50.800 --> 35:58.880 So we can say something like I'm going to enter into the expression and useful for when you need to 35:58.880 --> 36:05.120 answer questions about iOS 18 finds and returns the most relevant information from the documents in 36:05.120 --> 36:11.120 the vector database like features, update settings or changes in the new iOS version. 36:11.120 --> 36:14.880 Whenever you need information about iOS 18, you should always use this tool. 36:15.880 --> 36:16.240 All right. 36:16.280 --> 36:16.600 Great. 36:16.600 --> 36:17.880 So now let's close it. 36:19.200 --> 36:20.200 Let's save it. 36:20.400 --> 36:27.120 Now the best format for Llms, especially when they are doing semantic search through a vector databases, 36:27.360 --> 36:28.600 is actually marked down. 36:28.880 --> 36:31.730 So that's the most LM friendly Structure. 36:32.330 --> 36:35.570 As you probably remember, we're currently using a PDF. 36:35.850 --> 36:42.570 And while PDF is not the perfect format for our agent, it will still work completely fine in this setup. 36:43.330 --> 36:49.130 That said, we can improve the accuracy and relevance of the responses by taking one extra step. 36:49.130 --> 36:56.250 And that's, uh, specifying the pinecone namespace we created earlier in our vector store. 36:57.010 --> 37:04.370 So setting the namespace will help the agent stay focused on just the right chunk of data, and avoid 37:04.370 --> 37:08.690 mixing it with unrelated information if you are using multiple sources. 37:09.810 --> 37:16.170 So if you don't remember the name of the namespace you created, just go to your pinecone vector store 37:16.690 --> 37:17.890 and the namespace. 37:18.450 --> 37:24.250 And under namespaces you will see the name of the namespaces you created. 37:24.690 --> 37:26.490 So mine is rug agent. 37:28.290 --> 37:36.250 So now what we need to do we are we need to go to the pinecone vector node at option and add pinecone 37:36.250 --> 37:41.330 namespace and type the name of the namespace you created. 37:41.690 --> 37:47.650 As you can see, the partition the records in an index into namespaces. 37:48.770 --> 37:50.330 So mine is RAC agent. 37:53.410 --> 37:53.810 All right. 37:53.810 --> 37:54.130 Awesome. 37:54.130 --> 37:55.970 Now let's test our agent. 37:56.850 --> 37:58.330 For that let's click save. 37:59.530 --> 38:00.650 Let's open the chart. 38:01.490 --> 38:03.330 And now let's ask the same question. 38:15.850 --> 38:21.410 As you can see uh, now it is able to answer our question. 38:21.970 --> 38:26.770 So Image Playground in iOS 18 allows users to create and manipulate images through various tools and 38:26.810 --> 38:34.250 options, including creating images, generating images, saving sinking creations and so on. 38:34.290 --> 38:38.810 Now you can go to our document and check if it's true or not. 38:39.650 --> 38:41.010 Yes, you can see it's correct. 38:42.250 --> 38:42.650 All right. 38:42.690 --> 38:43.170 Awesome. 38:44.850 --> 38:45.930 So let's close it. 38:48.530 --> 38:48.890 All right. 38:48.890 --> 38:54.810 So now I'm going to show you how to share your agent with other people so they can start using it too. 38:55.090 --> 38:57.730 So first make sure your agent is set to active. 38:58.370 --> 39:01.930 Otherwise it would work for anyone outside your account. 39:01.930 --> 39:05.690 So just toggle it from inactive to active and you are good to go. 39:07.730 --> 39:08.730 And confirm it. 39:09.210 --> 39:18.890 Next click on your one chat message received trigger note and and then you have to make the chat publicly 39:19.130 --> 39:19.810 available. 39:21.970 --> 39:24.130 And here is the link you can share. 39:25.610 --> 39:32.540 So now you can copy this link and share it with clients team Members or embed in, uh, or embed it 39:32.580 --> 39:36.420 on a website so they will be able to open it in their browser. 39:36.660 --> 39:40.340 Ask questions and interact with your agent just like you do. 39:40.780 --> 39:43.500 So let's copy this link and see how it looks like. 39:45.500 --> 39:48.500 I'm going to paste it to my browser. 39:51.020 --> 39:51.700 And voila! 39:52.180 --> 39:56.180 As you can see, you can interact with your agent right here. 39:56.220 --> 39:58.180 You can of course change the initial message. 39:58.180 --> 40:03.340 So when you go to N810, you can modify the initial message right here. 40:04.660 --> 40:06.660 And you can make it look nicer. 40:06.660 --> 40:10.980 And I'm going to show you exactly how to do that in one of the upcoming lessons. 40:11.460 --> 40:12.820 So stay tuned for that. 40:13.940 --> 40:16.660 You have successfully built your first rack agent. 40:16.700 --> 40:23.940 You now have a fully working AI chatbot that can search your documents and answer questions using real 40:23.940 --> 40:24.460 data. 40:24.620 --> 40:26.060 So this is a huge milestone. 40:26.420 --> 40:29.260 So congrats again and see you in the next episode.