WEBVTT 00:00.040 --> 00:02.880 And I'm actually guessing it's going to be no surprise to you at all. 00:03.960 --> 00:09.840 Is the is the obvious next step to rag rag in traditional rag. 00:10.040 --> 00:11.600 We understand the process. 00:11.600 --> 00:18.760 You have chat with a user that goes to vector based retrieval that your code is written that is doing 00:18.960 --> 00:23.920 this, this lookup based on, uh, semantic search using vector similarity. 00:24.320 --> 00:27.760 I'm already using Pro Pro terms with you as if you've been doing this for years. 00:27.920 --> 00:34.880 Uh, and that is then sent to an LLM in order to get the LLM to respond to it, using the LLM to respond 00:34.880 --> 00:39.680 to this question with this extra context that is traditional Rag. 00:39.840 --> 00:47.400 So if you think about like bringing agents into this, the natural next step, uh, would wouldn't be 00:47.440 --> 00:48.120 necessarily this. 00:48.120 --> 00:49.400 This feels very linear. 00:49.440 --> 00:54.880 It feels very much as if you're doing you're writing code to do this vector based retrieval. 00:55.080 --> 01:01.000 And when you're thinking in the rag world, things are much more iterative and interactive. 01:01.000 --> 01:05.180 So what is the the agentic twist on this process? 01:05.180 --> 01:06.140 So here it is. 01:06.140 --> 01:10.420 The Agentic twist is to say, okay, we've got chat. 01:11.020 --> 01:13.060 The user messages us as before. 01:13.060 --> 01:14.180 What do we do with it? 01:14.300 --> 01:16.220 Well that goes to an agent. 01:16.220 --> 01:19.260 We it doesn't go yet to a vector based retrieval. 01:19.260 --> 01:20.820 It goes to an agent. 01:21.140 --> 01:24.740 And when I say it goes to an agent, what I mean is we're using an LLM. 01:24.900 --> 01:30.580 We're using an LLM not just for the response as we did before, but also the workflow, figuring out 01:30.580 --> 01:35.700 what to do next, what tools to use, how does it best answer those questions. 01:36.100 --> 01:43.340 This question, and one of those tools that we equip it with, is a tool that is able to do vector based 01:43.340 --> 01:44.260 retrieval. 01:44.260 --> 01:50.580 It's able it's given access to a vector database and said, hey, you can make vector queries of this 01:50.580 --> 01:55.940 vector database, and maybe you give it access to some of the different clever techniques around Rag 01:55.940 --> 01:57.020 that it could explore. 01:57.020 --> 01:57.940 If it wants to. 01:57.980 --> 02:04.270 It could, it could try it and then try a few other twists and turns as it tries to to find the relevant 02:04.270 --> 02:05.230 contexts. 02:05.230 --> 02:07.910 And hey, maybe we'll give it some other tools as well. 02:07.910 --> 02:10.790 You could imagine that maybe we just give it straight up SQL. 02:10.830 --> 02:18.310 Maybe if it turns out that there is our database table has has a country or a city destination column 02:18.310 --> 02:22.150 in it, and there's a table called Ticket prices, maybe it would be like, hey, I don't need to do 02:22.150 --> 02:25.390 all this fancy vector based stuff. 02:25.430 --> 02:32.150 I can just do like a SQL statement to select ticket prices where the city equals London. 02:32.150 --> 02:37.110 And so I'll write that SQL statement and I'll use my SQL tool to call the database and retrieve the 02:37.110 --> 02:39.790 information a more traditional kind of retrieval. 02:39.950 --> 02:48.350 And we give the the Lem the ability to control the workflow and do any of those things in order to find 02:48.390 --> 02:51.590 the best context for answering the question. 02:51.590 --> 02:55.190 And then that same Lem can then go ahead and answer the question. 02:55.190 --> 02:56.510 Now, of course, there are other ways of doing this. 02:56.510 --> 03:04.070 You can have a whole separate agent whose only job it is to to do the the retrieval part of it, and 03:04.070 --> 03:09.010 then it passes control on to another LM to actually answer the question. 03:09.050 --> 03:10.810 Of course, all of those things are possible. 03:11.130 --> 03:16.810 But in its simplest form, this is just a straight up comparison between traditional rag, which is 03:16.850 --> 03:19.650 a simple workflow, a linear workflow. 03:20.010 --> 03:21.090 Get the message. 03:21.210 --> 03:27.770 Look up relevant context, pass it to an LM versus a genetic rag, which is more iterative, where you 03:27.770 --> 03:34.170 let an LM decide how to do different kinds of techniques, to retrieve relevant context, to build up 03:34.170 --> 03:37.610 the relevant context in order to answer the question. 03:37.890 --> 03:41.570 So that is traditional rag versus a genetic rag. 03:41.610 --> 03:44.930 And no prizes for guessing which one we are going to be doing. 03:44.970 --> 03:45.330 Yeah. 03:45.330 --> 03:47.770 So it's going to be a genetic rag obviously. 03:48.130 --> 03:48.730 Okay. 03:48.810 --> 03:53.170 So today was it was very much a core expertise day. 03:53.210 --> 03:58.650 It might have felt like it was a lot of theory, but hopefully it was like applied theory very practical 03:58.650 --> 03:59.170 theory. 03:59.410 --> 04:04.850 But just to leave you with, with something to to do that isn't just slides. 04:04.930 --> 04:09.980 Uh, I wanted to spend a moment to mention super bass. 04:10.220 --> 04:16.620 So, uh, Nan comes with with lots of sort of, like, built in database stuff and some vector database 04:16.620 --> 04:17.580 things as well. 04:17.780 --> 04:24.620 But, but for this project, for this week, as we look to, to work with data and databases and even 04:24.620 --> 04:32.300 vector databases, I wanted to use a third party database that's incredibly popular and that's used 04:32.700 --> 04:38.380 by a lot of agencies and a lot of startups to midsize companies that are looking to quickly build products 04:38.380 --> 04:38.900 out there. 04:39.220 --> 04:41.940 Super bass, very, very popular. 04:41.980 --> 04:44.900 Definitely one that I thought it was worth us integrating with. 04:44.940 --> 04:47.540 And of course Ann has a super bass node. 04:47.780 --> 04:51.060 It had to because a super bass is really, really top. 04:51.460 --> 04:53.500 Okay so what what is super bass? 04:53.500 --> 05:01.140 Super bass is basically it uses it's it's a cloud version, a managed version of a type of database 05:01.140 --> 05:07.400 called Postgres, which is a fiendishly popular relational database for storing data. 05:07.600 --> 05:12.520 And super bass is so great because it's really easy to set up and use. 05:12.560 --> 05:18.280 It's very user friendly, it's very startup friendly and great for this kind of experience for you to 05:18.280 --> 05:19.200 get up and running. 05:19.200 --> 05:24.800 So I would like you to go to Super Bass, and I'd like you to spend a bit of time clicking around it. 05:24.800 --> 05:29.840 It also has a generous free tier of course, which is a must must have for this course. 05:30.000 --> 05:36.720 Uh, so I would like you to come on in here and just read about it, explore it. 05:36.840 --> 05:37.680 Uh, it does. 05:37.720 --> 05:43.680 They have a few different, um, service offerings, but the Postgres database is the big one. 05:43.720 --> 05:45.800 You can see they've got some pretty big name clients. 05:45.960 --> 05:53.520 Uh, and you'll see, amongst other things, that they store vector embeddings, which is the name for 05:53.520 --> 05:57.440 the vectors that come from embedding models, vector embeddings. 05:57.440 --> 06:02.280 They store vector embeddings in their database, which is just the kind of database that we're looking 06:02.280 --> 06:06.160 for in, in our project to build out a genetic rag. 06:06.160 --> 06:08.040 So that's your homework. 06:08.250 --> 06:09.330 Come on in here. 06:09.610 --> 06:10.530 Explore. 06:10.890 --> 06:12.290 Set up your account. 06:12.330 --> 06:13.410 Read about it. 06:13.410 --> 06:13.970 You can read. 06:13.970 --> 06:15.170 Read some of the docs. 06:15.210 --> 06:16.770 Get yourself somewhat familiar. 06:16.810 --> 06:22.170 See on the pricing page about the free plan and the paid plans, so that you get some sense of what 06:22.170 --> 06:22.810 we can do. 06:23.130 --> 06:26.730 Unlimited API requests 50,000 monthly active users. 06:26.730 --> 06:30.810 We're probably probably not going to get that with our project this week. 06:31.010 --> 06:34.730 Uh, and but you get a decent sense of what you can get. 06:34.770 --> 06:40.130 And then if you did want to pay more, then then this is what you would be able to do. 06:40.330 --> 06:43.250 Um, but that should be no need for this, this course. 06:43.610 --> 06:44.090 Okay. 06:44.130 --> 06:48.890 So I will leave you to do that, but also I will go back to the slides for the wrap. 06:48.930 --> 06:49.250 Okay. 06:49.290 --> 06:53.650 And I should mention that when you're setting up your super base account, I do believe it asks you 06:53.650 --> 06:55.970 to name your organization. 06:55.970 --> 07:01.490 And I recommend that what I did was I, I call my organization like donor research. 07:01.530 --> 07:04.450 Like my name and then research or education or whatever you want. 07:04.450 --> 07:11.030 Because this is you're using this as a personal, uh, playpen for, for your own research and education 07:11.030 --> 07:13.110 rather than for for a corporate purpose. 07:13.110 --> 07:14.070 So that's what I did. 07:14.270 --> 07:15.350 Uh, and different regions. 07:15.350 --> 07:20.830 You might have a slightly different sign up flow, but but I seem to remember it was it was super easy. 07:20.830 --> 07:24.510 And if you have any problems with that, of course, message me or look in the resources in case I've 07:24.510 --> 07:25.430 put any notes there. 07:25.710 --> 07:26.430 All right. 07:26.710 --> 07:34.510 So the next two days are where we are going to be building out a rag pipeline with a voice agent. 07:34.550 --> 07:41.550 The idea is we want to build an expert voice agent that could accelerate a business, because it could 07:41.550 --> 07:44.350 act as something that can handle more. 07:44.350 --> 07:50.390 Volumes of questions from people about your company, allow you to scale up, accelerate, because you 07:50.430 --> 07:56.230 are able to to do things in a streamlined way, because you can have experts working for you that are 07:56.230 --> 08:00.390 voice agents with complete knowledge of everything about your company. 08:00.510 --> 08:02.350 And that's that's what we're going to be building out. 08:02.350 --> 08:07.790 Just small steps, but leaving you with the ability to expand this so much, whether you're building 08:07.790 --> 08:11.120 that for your own company or whether it's something that you could build for others. 08:11.120 --> 08:16.080 You could offer that as a commercial product for other people that want to do this. 08:16.120 --> 08:20.240 It's a classic agentic use case and a rag use case. 08:20.440 --> 08:26.280 And so tomorrow we're going to be working with Superbase and integrating with Superbase for the first 08:26.280 --> 08:31.600 time and then taking data, turning it into vectors, and putting it into superbase. 08:31.600 --> 08:36.920 So it's going to be about data ingest and building automated pipes for that. 08:37.160 --> 08:40.360 And then that's going to be that's going to be great. 08:40.440 --> 08:46.360 And finally we're going to wrap up the week by building out our voice expert. 08:46.360 --> 08:47.560 That can then allow us. 08:47.600 --> 08:49.280 It's going to be just like what we did yesterday. 08:49.280 --> 08:53.640 We're going to have a voice agent that then calls a webhook running in. 08:53.680 --> 08:57.720 That's going to carry out a workflow to answer an expert question. 08:58.040 --> 08:59.760 So all of that is what's ahead. 09:00.040 --> 09:02.240 I hope you're excited, I certainly am. 09:02.360 --> 09:04.400 And with that, you're over halfway. 09:04.440 --> 09:11.440 53% of the way through this program, you're on on the path to being an Nanh Pro. 09:11.720 --> 09:12.480 See you tomorrow.