WEBVTT 00:00.040 --> 00:05.720 In this tutorial I will show you how to create a vector store in Supabase using Postgres. 00:05.920 --> 00:12.560 This is a powerful way to enable AI agents to remember key details from conversations, making interactions 00:12.560 --> 00:14.360 more natural and personalized. 00:15.160 --> 00:22.200 For example, in a WhatsApp AI agent for customer support I created for my own company which uses RAC 00:22.400 --> 00:27.720 and is trained on my company's knowledge base, memory allows it to go beyond just answering individual 00:27.720 --> 00:28.440 questions. 00:28.600 --> 00:33.640 It can remember the client's name, the previous inquiries, and the context of the conversation. 00:33.920 --> 00:39.920 This means clients can ask follow up questions without having to repeat themselves, creating a more 00:39.920 --> 00:41.320 human like experience. 00:41.840 --> 00:48.160 Additionally, the agent is also able to accept voice messages from clients, analyze them, and provide 00:48.160 --> 00:52.040 accurate responses, making the interaction even more engaging. 00:52.960 --> 00:56.840 To start, head over to spacex.com and log in. 00:56.880 --> 00:59.920 If you don't have an account yet, just register and create one. 01:01.080 --> 01:03.790 Once you are in, we need to create a new project. 01:04.350 --> 01:06.070 So click Create New project. 01:07.990 --> 01:10.470 Give it a name and set a strong password. 01:10.990 --> 01:13.830 You will want to save this password because you will need it later. 01:14.510 --> 01:18.830 Leave the default settings for the database region and hit Create New project. 01:19.190 --> 01:22.350 It will take a few moments for Supabase to set everything up. 01:23.310 --> 01:28.950 Now that our project is created, let's move on to setting up the database for our vector store. 01:29.390 --> 01:32.710 First we need to connect to our Supabase project in naan. 01:33.150 --> 01:39.630 So basically because we want to vectorize our data, for example knowledge base into our vector store, 01:39.830 --> 01:43.350 you have to drag and drop the vector store. 01:48.310 --> 01:51.750 And choose Add documents to vector store component. 01:52.630 --> 01:57.670 So now we have to set up credentials to connect to our super fast database. 01:58.270 --> 02:02.150 So I'm going to create new credentials from scratch. 02:04.830 --> 02:12.230 With new credential, then you have to go to our project in our Supabase account. 02:16.630 --> 02:17.310 Scroll down. 02:18.670 --> 02:20.310 You will see Project settings. 02:22.630 --> 02:24.510 Then click on API. 02:27.310 --> 02:30.070 And you should see Service Role. 02:30.110 --> 02:31.190 So click reveal. 02:32.870 --> 02:33.550 Copy it. 02:34.790 --> 02:35.390 Go back to. 02:37.990 --> 02:38.750 Paste it here. 02:41.110 --> 02:42.670 Then we need to provide host. 02:43.950 --> 02:48.670 So click on URL copy. 02:51.390 --> 02:54.710 And paste it here then save. 02:57.110 --> 02:59.350 And now again we will test our connection. 03:00.670 --> 03:03.190 So credential successfully created. 03:03.230 --> 03:08.900 So the next step is to create our vector store in database to store our data. 03:09.580 --> 03:11.300 So let's double click on this node. 03:13.180 --> 03:14.420 Then click on docs. 03:15.700 --> 03:18.780 It will text us to documentation. 03:19.420 --> 03:23.980 Scroll down click on Quick Start for setting up your vector store. 03:26.460 --> 03:27.500 Then scroll down. 03:28.380 --> 03:30.100 You will see a SQL code. 03:32.300 --> 03:33.140 Just copy it. 03:33.940 --> 03:35.940 Go back to your Supabase account. 03:37.660 --> 03:40.820 On the left hand side you should see SQL editor. 03:43.460 --> 03:46.780 Paste SQL code here and click run. 03:49.420 --> 03:52.420 This process will create your vector store automatically. 03:53.980 --> 04:00.220 Now, in order to check if we successfully created our vector store, let's click on Table Editor. 04:02.420 --> 04:06.020 And as you can see we have a table named documents. 04:07.740 --> 04:13.660 So here we will store our vectorized knowledge base in the form of vectors. 04:14.180 --> 04:20.140 Now in our workflow, when you click on supervised vectors or nodes. 04:22.620 --> 04:27.500 You will see that we are able to choose our vector store from the list. 04:30.900 --> 04:37.740 Then we have to add option and leave default option match documents as it is. 04:40.980 --> 04:45.260 The next step is to choose embeddings model. 04:45.620 --> 04:48.500 So let's choose OpenAI. 04:53.340 --> 04:59.100 And this model text embedding model is is really cost effective. 04:59.580 --> 05:01.500 So let's choose this one. 05:07.260 --> 05:14.490 So change the type of data to binary so it will process binary data from the previous step. 05:17.410 --> 05:23.890 Then leave mode as it is and data format automatically detect by Mime type. 05:26.770 --> 05:30.050 The next step is to split larger text into smaller chunks. 05:30.490 --> 05:36.850 This is important because AI models like OpenAI's embeddings perform better when processing smaller 05:36.850 --> 05:42.210 segments of text, ensuring more accurate embeddings and efficient processing. 05:42.970 --> 05:51.410 So in this example, uh, I'm going to select recursive character text splitter. 05:52.330 --> 05:58.730 This option is ideal because it recursively breaks down the text into smaller chunks while maintaining 05:58.730 --> 05:59.450 context. 06:00.010 --> 06:06.970 And this ensures that no important details are lost, uh, during the splitting process, and the chunks 06:06.970 --> 06:09.490 remain manageable for embedding generation. 06:12.450 --> 06:19.720 So usually I'm using, uh, 200 character overlap because it works perfectly for shorter documents and 06:20.360 --> 06:22.760 keep the context intact between chunks. 06:24.560 --> 06:26.240 So let's type 200. 06:28.280 --> 06:30.200 Start our vectorization process. 06:30.520 --> 06:37.840 We need a trigger to initiate the workflow and provide data that can be vectorized in supabase. 06:38.680 --> 06:48.440 So in this example, as you can see, I'm using a Google Drive trigger to download a file of my company's 06:48.440 --> 06:48.920 data. 06:48.960 --> 06:50.080 My company's data. 06:50.560 --> 06:57.720 And and this file serves as the source data which is then processed and vectorized into space. 07:01.240 --> 07:01.720 All right. 07:01.720 --> 07:04.360 So let's move on to the second part of the workflow. 07:08.120 --> 07:15.120 So in this section we will focus on retrieving information from our vector store and connecting it to 07:15.160 --> 07:18.440 our AI agent to generate responses for the user. 07:20.200 --> 07:23.080 First, add the AI agent node. 07:23.880 --> 07:29.920 So this node acts as a trigger for this part of the workflow. 07:30.360 --> 07:34.640 So it starts the process when a user sends a message. 07:35.320 --> 07:41.680 Ensuring the workflow receives their query and manages the overall logic of the interaction. 07:42.400 --> 07:54.880 So drag this node and canvas and configure it to listen for input messages and set it up. 07:54.920 --> 08:01.840 Set it up to with dynamic expressions for processing user data and session context. 08:04.520 --> 08:08.000 So next uh, add the Postgres chat memory node. 08:08.520 --> 08:15.160 So this ensures that the agent can retrieve and use previous conversation history. 08:15.560 --> 08:23.750 So this way the agent remembers the user's name, previous inquiries, or any other contextual information. 08:24.310 --> 08:31.790 So we need to set up the database connection to your Postgres instance and map the memory fields like 08:32.150 --> 08:33.910 username and queries. 08:40.550 --> 08:45.910 So here we have to set up credentials to connect with Supabase. 08:46.990 --> 08:48.350 So it can be quite tricky. 08:48.350 --> 08:51.590 But uh, I will guide you step by step. 08:53.910 --> 08:55.870 So let's create credentials from scratch. 08:59.230 --> 09:10.350 So to configure the Postgres connection go to your database dashboard and locate Project Settings. 09:11.830 --> 09:13.350 Then click on database. 09:14.990 --> 09:18.230 And as you can see we have all the necessary details. 09:19.100 --> 09:24.580 So we have to just copy it and paste into the fields. 09:25.180 --> 09:30.900 So let's copy the host, go back to an workflow. 09:32.100 --> 09:32.860 Paste it here. 09:35.140 --> 09:38.220 Do the same with a port. 09:42.740 --> 09:43.620 Scroll down here. 09:50.100 --> 09:51.780 And then the user. 10:01.420 --> 10:08.260 And and here we have to enter the password associated with our Postgres user account. 10:13.300 --> 10:14.180 Then click save. 10:16.220 --> 10:20.620 Once all the fields are filled, click on Test Connection to verify the setup. 10:21.140 --> 10:25.300 and if it's successful, your node is now connected to the Postgres database. 10:33.740 --> 10:41.580 After that, before we add the supervised vector store node, we need to set up the vector store tool 10:41.580 --> 10:42.020 node. 10:42.460 --> 10:49.780 So this node is responsible for interacting with your knowledge base stored in the vector store and 10:49.780 --> 10:52.460 preparing relevant information for your AI agent. 10:54.460 --> 11:00.500 So by clicking on this plus button let's search Vector Store tool. 11:06.340 --> 11:12.060 And you have to provide a name then description. 11:12.060 --> 11:18.620 So basically it's an instruction to AI agent what to do. 11:18.660 --> 11:20.620 So in this example. 11:23.170 --> 11:28.610 Uh, I provided a description of this tool to query the adaptive II knowledge base. 11:31.970 --> 11:34.210 Now add the AI model. 11:36.330 --> 11:39.050 So I'm gonna choose open AI chat model. 11:42.650 --> 11:49.730 You have to create a credential to connect with open AI account to to use your API keys. 11:51.610 --> 11:57.930 I really like GPT for omni model because, uh, it's really cost effective and works really well for 11:57.930 --> 11:59.010 this purpose. 12:00.410 --> 12:02.690 Now we have to choose our chat model. 12:04.410 --> 12:06.930 Let's select the same model as previously. 12:06.930 --> 12:09.610 So ChatGPT for me. 12:12.930 --> 12:22.010 And finally we can add our supervised vector store node to retrieve embeddings from the vector database. 12:28.130 --> 12:28.450 So. 12:30.570 --> 12:31.010 Sir. 12:32.890 --> 12:38.850 And now we have to choose our table to retrieve information. 12:38.850 --> 12:41.530 So from the list. 12:45.490 --> 12:46.690 Select documents. 12:51.290 --> 12:56.690 And now we have to choose our embedding model. 12:56.690 --> 13:04.450 So this is an important step because we have to select the same embedding model as in the first part 13:04.450 --> 13:05.530 of our workflow. 13:08.290 --> 13:12.770 So open the I text embedding three model. 13:22.970 --> 13:27.650 So this is how to create a supervised vector store in Nan.