WEBVTT 00:00.080 --> 00:02.960 All right, here we are, our super bass project rag. 00:02.960 --> 00:03.800 I went and got a coffee. 00:03.800 --> 00:04.840 I came back and it's done. 00:04.840 --> 00:05.920 It takes a few minutes. 00:06.360 --> 00:09.840 So let's just look around this navigation as we're looking into our project. 00:09.880 --> 00:14.960 Of course, the way you get here if you ever get lost is if you're on your projects page, you simply 00:14.960 --> 00:16.440 come into rag and back. 00:16.440 --> 00:17.560 We are here again. 00:17.560 --> 00:20.600 And looking at the nav here, we can be looking at tables. 00:20.600 --> 00:26.400 We can looking at something called SQL, SQL which you probably know is how you, you you ask questions, 00:26.400 --> 00:32.200 you query your database, the database itself, uh, and then a bunch of other stuff. 00:32.440 --> 00:35.920 If we look at the database itself, loading tables, we have no tables. 00:35.920 --> 00:37.760 It's an empty database. 00:38.080 --> 00:38.600 Okay. 00:38.640 --> 00:39.000 Back. 00:39.000 --> 00:41.920 We go back we go to the project overview. 00:42.160 --> 00:43.000 Here we are. 00:43.200 --> 00:48.400 And now is the one thing I told you I was going to do something that's kind of Cody. 00:48.400 --> 00:53.160 And so for those of you that are engineers this is going to be like, sure, this is easy for those 00:53.160 --> 00:55.880 of you that aren't this is going to be like, Holy cow, what is this? 00:56.120 --> 01:02.720 And and I'm here to tell you that it honestly, it's very cookie cutter stuff. 01:02.880 --> 01:08.230 Whilst you do need to run some code at this point, which you will find in the in the course resources 01:08.270 --> 01:09.830 tell you exactly what to do. 01:10.030 --> 01:13.070 This is a very pedestrian activity. 01:13.350 --> 01:14.550 And here's the thing. 01:14.870 --> 01:20.630 You could ask ChatGPT to write this for you with some instructions, and it would do it in a heartbeat. 01:20.830 --> 01:24.710 And in fact, as like a just a dark secret that's just between you and me. 01:24.750 --> 01:25.910 Don't tell anybody else. 01:25.910 --> 01:29.150 But I got ChatGPT to generate this for me because I hate SQL. 01:29.150 --> 01:31.590 I always seem to make mistakes with it, so I didn't want to do that. 01:31.590 --> 01:33.310 I just asked ChatGPT to generate it. 01:33.310 --> 01:36.070 I told it exactly what it needed to do and it did it right away. 01:36.470 --> 01:42.190 You can use that or code or whatever you like to use, but when you're ready, I need you now to go 01:42.190 --> 01:48.670 to the SQL editor, which is where you get to run code should you wish, by pasting it in there and 01:48.670 --> 01:50.470 pressing that run button down there. 01:50.510 --> 01:51.910 And that's what we're going to do next. 01:51.910 --> 01:58.390 So the reason we're doing this is because the tables that come with the super base database aren't, 01:58.390 --> 02:02.390 by default, set up to be exactly the format that Nam would like. 02:02.390 --> 02:06.710 And it's not like there's a there's a sort of out of the box way to say, hey, make me the tables and 02:06.750 --> 02:07.310 expects. 02:07.310 --> 02:11.910 So we have to create the database in that particular format so that this will work well. 02:11.950 --> 02:15.750 And so if I paste in this SQL, this is the SQL that you should be able to copy and paste also from 02:15.750 --> 02:16.510 the resources. 02:16.510 --> 02:21.790 And let me just quickly say at a high level, because there is one janky thing about this one janky 02:21.830 --> 02:22.150 thing. 02:22.350 --> 02:27.510 Uh, so this basically says, I want you to create a table called knowledgebase that is going to be 02:27.510 --> 02:28.630 the database table. 02:28.630 --> 02:29.630 It's going to create. 02:29.790 --> 02:35.310 And I want every row on that table to have an ID content, which is going to be the text that we just 02:35.310 --> 02:41.790 built in metadata, which is going to be the metadata including the category and then the vector. 02:42.070 --> 02:44.030 The embedding vector. 02:44.030 --> 02:46.350 And this is the janky thing that I'm going to come back to. 02:46.790 --> 02:51.950 And then I need to make a single a function in the database called a match documents. 02:51.950 --> 02:55.230 And this is a function that expects this is a function. 02:55.230 --> 02:59.350 It's going to want to call when it wants to retrieve relevant context. 02:59.350 --> 03:03.910 And as I say, if you go with like like some database setups, this will all be automatically built 03:03.910 --> 03:06.350 and will be exactly what Nan is expecting. 03:06.350 --> 03:09.510 And we only had to do this because we're going pro with Supabase. 03:09.790 --> 03:12.790 So we're creating this function, match documents. 03:12.790 --> 03:14.830 We're saying that it should return some information. 03:14.940 --> 03:17.220 and this is what the function needs to look like. 03:17.260 --> 03:24.580 And basically it's because when you when you do the natural calculation in the base table with this, 03:24.580 --> 03:29.020 this Postgres function, it gives something slightly different to what Nam would like. 03:29.060 --> 03:34.580 And you have to take one minus that calculation in order to get just the right number for super base. 03:34.580 --> 03:35.780 So this stuff doesn't really matter. 03:35.780 --> 03:41.580 If you wanted to generate this for you, you would just simply say the ChatGPT hey, can you make me 03:41.940 --> 03:42.420 something? 03:42.420 --> 03:49.060 I can run in super base in Postgres, which would allow me to set up a table called Knowledge Base that 03:49.060 --> 03:55.180 would be in the right format with content and metadata, and a vector that Nam would expect, and it 03:55.180 --> 03:58.100 will generate something just like this, as it did for me. 03:58.420 --> 04:01.220 Okay, now let me explain the one janky thing. 04:01.220 --> 04:02.060 So here's the thing. 04:02.380 --> 04:07.340 People that are relatively new to Rag sometimes think that the the magic in rag is something to do with 04:07.340 --> 04:08.940 what's in the vector database. 04:08.940 --> 04:14.900 These vectors and and whilst it's true that magic is in the vectors, the vector database, it's just 04:14.900 --> 04:18.740 an efficient database that can deal with with, with similarity calculations. 04:18.740 --> 04:22.260 The thing that's doing the magic is the embedding model. 04:22.260 --> 04:28.060 It's the LLM that's able to take text and turn it into a vector, and that's outside Postgres. 04:28.060 --> 04:32.820 We're about to set that up in N810, and that's going to come up with a vector and pass it in. 04:32.980 --> 04:38.900 And what Postgres and Superbase is able to do well is store that vector. 04:38.900 --> 04:42.100 But but this database doesn't actually create the vector. 04:42.500 --> 04:47.780 And that means that there is one important thing that needs to be set up when you create the database. 04:47.780 --> 04:54.180 And it comes down to this, this idea of, of dimensions, it's the question is how many dimensions 04:54.180 --> 04:55.740 are the vectors going to have. 04:55.780 --> 04:59.380 Because this database needs to be set up to expect that. 04:59.380 --> 05:03.700 And that is a is a is a feature of the embedding model that we choose. 05:03.740 --> 05:07.580 As I say, you can imagine one that just comes up with three numbers that you can imagine. 05:07.580 --> 05:09.220 Some come up with thousands. 05:09.220 --> 05:11.580 And so you have to pick a model. 05:11.740 --> 05:17.420 And then that you have to make sure that the number of dimensions that the database is expecting matches 05:17.420 --> 05:18.540 your embedding model. 05:18.740 --> 05:26.940 And in our case, we're going to pick a really standard typical first choice, which is to use a model 05:26.940 --> 05:30.340 that called called OpenAI embedding small. 05:30.340 --> 05:38.020 And it's a great model and everyone uses it, and it happens to have 1536 dimensions. 05:38.020 --> 05:42.380 And that number one, five, three, six needs to go in these two places here. 05:42.420 --> 05:44.060 And that is a bit janky. 05:44.060 --> 05:45.900 I do agree that seems like what. 05:45.940 --> 05:51.300 So if you put one, five, three, seven one off, you're going to get a weird error at some point about 05:51.300 --> 05:52.940 mismatching dimensions. 05:52.940 --> 05:53.940 And that's no good. 05:53.980 --> 06:00.060 So that has to be right now OpenAI embeddings model does have a charge associated with it, but it's 06:00.060 --> 06:01.180 really tiny. 06:01.220 --> 06:03.420 It's really tiny for huge numbers of documents. 06:03.420 --> 06:05.780 And so it's super tiny for 60 documents. 06:05.820 --> 06:10.260 Nonetheless, I know that that OpenAI needs an upfront payment and you might not want to use that. 06:10.260 --> 06:14.940 Gemini has a great embeddings model two, and if you know what you're doing with open source, of course 06:14.940 --> 06:16.700 the open source models are free. 06:17.180 --> 06:21.140 And if you do choose to go in that different direction, the only thing you need to know is once you've 06:21.140 --> 06:26.140 picked your embedding model, you may have to come back and rerun this with a different number. 06:26.140 --> 06:30.490 In here you can run this, this, this recreates the table if it already exists. 06:30.490 --> 06:32.010 You can run this as many times as you want. 06:32.050 --> 06:38.690 Just make sure that the last time you run it that the number of dimensions that you specify here matches 06:38.690 --> 06:40.650 the embedding model that you pick. 06:40.690 --> 06:48.050 And if you stick with the plan and you pick OpenAI's model, then 1536 is the perfect number to have 06:48.050 --> 06:48.450 here. 06:48.450 --> 06:50.490 So I appreciate that. 06:50.530 --> 06:52.930 It's a bit weird to to to have it like this. 06:52.970 --> 06:58.290 It seems like a bit of voodoo, but but uh, yeah, you need to make sure that the dimensions in your 06:58.290 --> 07:03.810 embedding model matches the number of dimensions that you tell super base it needs to create in this 07:03.810 --> 07:04.690 database table. 07:05.010 --> 07:07.650 All right, that's a long explanation. 07:07.690 --> 07:08.570 Time to run this. 07:08.610 --> 07:10.570 Ah, but wait, there is one more thing we need to do. 07:10.610 --> 07:11.770 Just before we run this script. 07:11.770 --> 07:18.970 We have to just tell Super Base that we're using vector related stuff with this database. 07:18.970 --> 07:25.610 And the way we do that is you go over here on the left to the database section, and you go to extensions 07:25.770 --> 07:26.730 right here. 07:26.730 --> 07:30.010 And you see all these extensions here, type vector. 07:30.170 --> 07:31.330 And this comes up. 07:31.330 --> 07:35.120 And then you turn that on and here we go. 07:35.160 --> 07:36.520 Enable extension. 07:36.760 --> 07:37.640 Let it think. 07:38.160 --> 07:38.880 And it's done. 07:38.880 --> 07:42.200 Extension vector is now enabled. 07:42.400 --> 07:45.480 And now come back to the SQL editor. 07:45.600 --> 07:46.320 Here we go. 07:46.360 --> 07:49.320 You see the arrow I got down there because I didn't do it yet. 07:49.400 --> 07:53.160 Uh, and now we should be able to run and get a better result. 07:53.200 --> 07:54.120 Okay, here we go. 07:54.120 --> 07:54.680 Run! 07:55.000 --> 07:55.760 Running. 07:55.760 --> 07:59.160 It says success. 07:59.160 --> 07:59.840 No rows returned. 07:59.840 --> 08:00.480 That's what we wanted. 08:00.480 --> 08:02.120 We didn't want any rows returned. 08:02.320 --> 08:04.360 And now to check that it's really been successful. 08:04.360 --> 08:06.040 Like it says we go to database. 08:06.400 --> 08:08.560 It says it said for a moment it said no tables. 08:08.560 --> 08:09.200 But there it is. 08:09.200 --> 08:11.720 There is our knowledge base database we've set up. 08:11.960 --> 08:16.320 It has an ID, it has content that's going to be the stuff. 08:16.360 --> 08:21.720 It has metadata and it has embedding, which is our vector. 08:21.920 --> 08:24.280 This is where it's going to go. 08:24.480 --> 08:26.040 So we've done it. 08:26.040 --> 08:29.040 We've set up a super base database. 08:29.040 --> 08:37.360 It's now time for us to go to back to N and to hook it all up and then populate our database, do the 08:37.400 --> 08:39.280 load part of ETL.