WEBVTT 00:00.120 --> 00:00.440 Well. 00:00.440 --> 00:01.120 Hello there. 00:01.120 --> 00:02.920 Today is a blue day. 00:02.960 --> 00:07.560 Today is completing the Business Commercial project as part of week two. 00:07.840 --> 00:11.200 All about accelerating your business or your client's business. 00:11.440 --> 00:12.840 I put it to you. 00:12.840 --> 00:18.200 I believe that today is going to be a day that is going to be relatively low on challenge and relatively 00:18.200 --> 00:21.440 high on satisfaction, but you'll be the judge of that. 00:21.440 --> 00:22.080 Let's see. 00:22.120 --> 00:29.000 Today we are wrapping up our expert product for being an expert for your client. 00:29.000 --> 00:34.800 And we're going to be building our rag agent and our voice agent, of course, bringing back the return 00:34.840 --> 00:36.680 of 11 labs and voice. 00:37.080 --> 00:42.600 And of course, I have to begin one more time with a recap on rag, and this will be the last one for 00:42.600 --> 00:43.000 rag. 00:43.200 --> 00:45.640 But it's always, always good stuff to go through. 00:45.640 --> 00:49.840 There's a couple of times repetition is the best way to learn, but you're pretty familiar with this 00:49.840 --> 00:50.520 stuff now. 00:50.560 --> 00:53.080 The user asks a chat question. 00:53.080 --> 00:59.640 It comes into our code, perhaps running in n810 and the big idea behind rag. 00:59.640 --> 01:04.950 We want to try and look up some useful information, and the way we do it is we use a different kind 01:04.950 --> 01:12.710 of alarm called an encoder, or an embedding model, or a vector embedding model, or, uh, any, any 01:12.710 --> 01:13.550 of those things. 01:13.710 --> 01:18.590 And it takes the text and it turns it into a set of numbers that we call a vector. 01:18.790 --> 01:28.110 And if you're using the very popular OpenAI embedding small, then it comes back with 1536 numbers. 01:28.550 --> 01:29.510 Correct me if I'm wrong. 01:29.710 --> 01:36.390 Uh, and, uh, you can think of that as being like a point in 1536 dimensional space. 01:36.550 --> 01:42.790 And it's done such that you can then go to all of your data sitting in a type of database that we sometimes 01:42.790 --> 01:43.790 call a knowledge base. 01:43.950 --> 01:45.070 And you have indexed it. 01:45.070 --> 01:50.550 You've got all of the bits of data that have also already had vectors calculated for them. 01:50.550 --> 01:51.590 That's what we did yesterday. 01:51.750 --> 01:55.950 So we've got all of these 1536 dimensional vectors. 01:56.150 --> 01:57.990 And we can find the ones that are closest. 01:57.990 --> 02:03.830 And that's likely to give us some material that is going to be relevant to answering the question and 02:03.830 --> 02:08.350 we take that material, we take the original text associated with it, and we send that to the LLM. 02:08.350 --> 02:14.870 And back comes a knowledgeable response, giving the illusion that the LLM itself is, is knows about 02:14.870 --> 02:19.430 everything in the database, because whatever question you ask, we bring back relevant context for 02:19.470 --> 02:19.990 that. 02:19.990 --> 02:24.150 And we learned about chunking and the fact that you typically take your documents and divide them up. 02:24.150 --> 02:27.910 But as it happened, none of our of our documents were less than 1000 characters. 02:27.910 --> 02:33.550 So we ended up with exactly one chunk per row in the Google sheet that we used as our source. 02:33.550 --> 02:35.070 And there are two different phases of Rag. 02:35.110 --> 02:40.950 There's the data ingest, which is what we did yesterday, and data ingest is where you have some source 02:40.950 --> 02:41.670 of data. 02:41.670 --> 02:44.510 And you first of all, extract your data from it. 02:44.510 --> 02:47.230 You transform it in our case with the field mapping node. 02:47.270 --> 02:53.430 And then and then you can chunk it and vectorize it or form vector embeddings from it. 02:53.430 --> 02:56.870 And then you load that into your vector store. 02:56.870 --> 03:03.910 In our case we used a Postgres database in Superbase, the managed Postgres service. 03:04.070 --> 03:08.300 And this is something that can run once as a way to load in your vector store. 03:08.300 --> 03:12.140 But it's common to have this running as kind of live data pipes. 03:12.140 --> 03:14.300 And you could do that in through a few different ways. 03:14.300 --> 03:19.780 You could do it by by having it run on a schedule so that it brings in things on some schedule or the 03:19.780 --> 03:21.420 thing that you could do, should you wish. 03:21.420 --> 03:26.860 And I will really equip you to do it next week is have something that like, listens on a shared drive, 03:26.860 --> 03:32.460 like a Google folder, and then when a new document is dropped in there, it automatically kicks this 03:32.460 --> 03:37.540 whole process off, loads in that document, and then puts all of that document and all of its chunks 03:37.540 --> 03:38.820 into your vector store. 03:38.860 --> 03:40.420 That would be that would be really cool. 03:40.420 --> 03:45.260 And that's the kind of thing you'd actually have running live in production, constantly keeping your 03:45.260 --> 03:46.940 vector store up to date. 03:46.940 --> 03:52.140 And of course, the other side of this is the thing that actually answers the user's question. 03:52.180 --> 03:58.700 The user asks a question, we have an AI agent that can answer it, and the agent has an LLM powering 03:58.700 --> 04:03.620 it that that is used both to actually give the answer, but also to orchestrate how to answer it. 04:03.620 --> 04:08.940 And as part of doing that, it has access to a tool, and that tool is a tool that allows it to vectorize 04:08.940 --> 04:13.940 the question that's being asked, and look it up in the vector store to find relevant content. 04:13.940 --> 04:16.820 And this is, of course, what we call a genetic rag. 04:16.820 --> 04:20.620 And you could also equip it with other tools to dig through the data as well. 04:20.860 --> 04:24.300 And so that is the second phase, which is what we'll be focusing on today. 04:24.340 --> 04:30.660 And as well, you know, the the business challenge that I set for this week is this kind of classic 04:30.900 --> 04:38.020 business, uh, commercial opportunity for for genetic AI, for generative AI, which is an expert question 04:38.020 --> 04:45.100 answerer that's able to to answer questions as if the LLM knows all about your business or in our case, 04:45.100 --> 04:51.300 all about the products that your client offers and can answer any kind of question about these products. 04:51.540 --> 04:57.220 And as I say, so classic, it's so clear to see how you can use this to accelerate a business, to 04:57.220 --> 05:02.620 be able to have workers effectively at the business that are able to do the more manual, more menial 05:02.660 --> 05:03.420 tasks. 05:03.740 --> 05:10.020 By manual, I mean the more sort of manually intensive Tasks that should be ripe for this kind of solution. 05:10.260 --> 05:13.380 And for for our solution, we're using Supabase. 05:13.380 --> 05:20.700 It is a provider, a cloud provider of Postgres, which is a very common, scalable relational database 05:20.700 --> 05:23.980 used by, by by, by many enterprises. 05:24.180 --> 05:29.900 It's, uh, we're using some sort of advanced approaches, but we have to we have to copy and paste 05:29.900 --> 05:32.460 in that code, that slightly janky code. 05:32.460 --> 05:37.580 And I promised you you could just generate that with ChatGPT if you had to do it on your own, and it 05:37.580 --> 05:38.260 would do it for you. 05:38.260 --> 05:41.700 It could explain it to you, it can fix it if there are problems. 05:41.700 --> 05:44.700 So whilst this is meant to be a low code, it's not a no code. 05:44.700 --> 05:46.100 It's a low code course. 05:46.100 --> 05:48.780 So I hopefully don't do too much coding. 05:48.780 --> 05:52.860 That was one moment when we looked at it, but it's it's cookie cutter stuff. 05:52.860 --> 05:54.660 You can use it anywhere. 05:55.020 --> 05:58.820 And we worked on the data ingest side of the house yesterday. 05:58.980 --> 06:01.220 We did some interesting work to do that, that field mapping. 06:01.220 --> 06:06.340 If you remember, transforming data from one format in the Google Sheet to the format with a content 06:06.340 --> 06:10.170 and category that we wanted to put in our vector data store. 06:10.410 --> 06:11.570 And that is done. 06:11.570 --> 06:16.730 And we've loaded up our rag, our knowledge base in our Rag project in Super Base. 06:16.730 --> 06:22.010 It has 60 rows, which all have vectors with 1536 dimensions. 06:22.010 --> 06:22.970 I hope I'm getting that right. 06:23.050 --> 06:25.690 Uh, and we've got all that in there. 06:25.770 --> 06:27.650 Remember those dimensions? 06:27.650 --> 06:32.210 The clever part comes from the encoder, from the OpenAI embeddings. 06:32.210 --> 06:33.450 Small model. 06:33.690 --> 06:35.050 That's what did it. 06:35.090 --> 06:41.410 And our mission for today, of course, is to add the questioning, questioning question answering part 06:41.410 --> 06:45.330 of this and then add a voice agent and then we will declare victory. 06:45.330 --> 06:50.090 Well, here we are back in town, and it's time for us to take a final look at this beautiful ingest 06:50.090 --> 06:53.730 pipeline and then make sure that you've saved it. 06:53.850 --> 07:01.330 And then go back to this screen and it's time to create a new workflow. 07:01.410 --> 07:06.650 This is going to be the home of our question answerer, our Agentic AI. 07:07.050 --> 07:13.480 And we're going to start this time by having a first step that is going to be a good old on chat message. 07:13.880 --> 07:16.120 Our original, our original and favorite. 07:16.200 --> 07:17.000 Uh, there it is. 07:17.000 --> 07:18.120 When chat message is received. 07:18.120 --> 07:18.720 We know this one. 07:18.720 --> 07:20.280 Well, that's how we will begin. 07:20.440 --> 07:21.360 And you know the deal. 07:21.360 --> 07:24.160 The next thing we do is we hit tab to add a node. 07:24.160 --> 07:29.920 We want to add an AI node AI agent and then press escape. 07:29.920 --> 07:31.000 And here it is. 07:31.200 --> 07:36.400 And for once it's good to leave that uh, if I double click, it's fine to keep this connected chat 07:36.440 --> 07:38.040 trigger node and JSON input. 07:38.040 --> 07:39.120 That's exactly what we want. 07:39.160 --> 07:41.960 You now have a better idea about what that means and why it says that. 07:42.080 --> 07:45.360 Uh, so, uh, let's leave that go to chat model. 07:45.400 --> 07:46.960 I guess we'll go with Gemini again. 07:46.960 --> 07:48.360 Gemini chat model. 07:48.360 --> 07:50.120 Of course, you could pick whatever chat model you want. 07:50.120 --> 07:52.240 Use open router if you like that. 07:52.400 --> 07:59.880 Uh, and we will go with models slash Gemini three flashcards. 08:00.240 --> 08:01.080 Preview. 08:01.120 --> 08:02.120 There it is. 08:02.600 --> 08:05.880 And, uh, that's the model we're picking here. 08:06.040 --> 08:07.960 Memory will just use a simple memory. 08:08.400 --> 08:11.520 And it's time for us to add in the tool.