WEBVTT 00:00:00.000 --> 00:00:03.040 Imagine an AI chatbot, you know, brilliant at 00:00:03.040 --> 00:00:05.419 general conversation. But then you ask it about 00:00:05.419 --> 00:00:07.599 your company's specific policies or maybe the 00:00:07.599 --> 00:00:10.060 nuances of your industry. And he either guesses 00:00:10.060 --> 00:00:12.519 or maybe worse, just kind of makes things up. 00:00:12.960 --> 00:00:15.919 What if you could give that AI a custom brain, 00:00:16.179 --> 00:00:20.420 one filled with your verified knowledge? Welcome 00:00:20.420 --> 00:00:23.140 to the deep dive. Today, we're diving deep into 00:00:23.140 --> 00:00:25.699 exactly that idea. Retrieval Augmented Generation, 00:00:26.019 --> 00:00:28.899 or RA for short. And honestly, it's simpler than 00:00:28.899 --> 00:00:30.780 it sounds, but it fundamentally changes how we 00:00:30.780 --> 00:00:33.439 can trust and actually use AI. We're going to 00:00:33.439 --> 00:00:36.320 discover what RA is, how these powerful Mind 00:00:36.320 --> 00:00:38.700 Palace databases make it possible. We'll break 00:00:38.700 --> 00:00:40.859 down the two crucial parts of any RA system. 00:00:41.100 --> 00:00:42.700 Then yeah, we'll walk you through building one, 00:00:42.759 --> 00:00:45.200 step by step, no code needed. You'll even learn 00:00:45.200 --> 00:00:46.979 how to give your AI a memory, make it remember 00:00:46.979 --> 00:00:49.619 stuff, and power its new brain efficiently. It's 00:00:49.619 --> 00:00:51.530 really about turning a smart... chatbot into 00:00:51.530 --> 00:00:54.030 a true expert on your stuff. Okay, let's unpack 00:00:54.030 --> 00:00:56.130 this. So let's start right at the heart of it. 00:00:56.229 --> 00:00:59.850 What is ARAG? Retrieval Augmented Generation. 00:01:00.829 --> 00:01:03.770 In plain English, it just means your AI does 00:01:03.770 --> 00:01:05.549 its own research before it speaks. Think of it 00:01:05.549 --> 00:01:07.230 like an open book test. When you ask a question, 00:01:07.290 --> 00:01:11.250 maybe how many feet are in a mile? ARAG system 00:01:11.250 --> 00:01:13.269 doesn't just pull from its huge general training 00:01:13.269 --> 00:01:16.349 data. No, it first retrieves specific accurate 00:01:16.349 --> 00:01:20.109 info like 5 ,280 feet from a source you've given 00:01:20.109 --> 00:01:22.900 it. Only then does it use that verified fact 00:01:22.900 --> 00:01:26.219 to augment or, well, significantly improve its 00:01:26.219 --> 00:01:28.879 answer. This is what transforms a generic AI 00:01:28.879 --> 00:01:31.340 into one that genuinely knows your company docs, 00:01:31.560 --> 00:01:33.819 your support tickets, maybe your internal policies. 00:01:34.060 --> 00:01:36.459 Yeah, it changes the AI from just a generalist 00:01:36.459 --> 00:01:39.359 into a trusted specialist. OK, so the core idea 00:01:39.359 --> 00:01:41.739 is that the AI isn't just generating from like 00:01:41.739 --> 00:01:44.200 its initial programming. It's actively seeking 00:01:44.200 --> 00:01:46.640 out and pulling in external facts first. Exactly. 00:01:46.739 --> 00:01:48.859 It finds the facts, then generates with confidence. 00:01:49.390 --> 00:01:52.609 But where does this AI store all the specific 00:01:52.609 --> 00:01:55.430 knowledge so it can find it so fast? That's where 00:01:55.430 --> 00:01:57.250 the vector database comes in. It's truly the 00:01:57.250 --> 00:02:01.010 brain of a rag agent. Imagine this vast... multi 00:02:01.010 --> 00:02:03.269 -dimensional space kind of like a galaxy right 00:02:03.269 --> 00:02:05.829 where every star is a piece of information from 00:02:05.829 --> 00:02:08.650 your documents each chunk of text becomes a single 00:02:08.650 --> 00:02:10.310 point of light in this galaxy that's what we 00:02:10.310 --> 00:02:12.750 call a vector and the really profound insight 00:02:12.750 --> 00:02:16.310 here is semantic similarity this database doesn't 00:02:16.310 --> 00:02:18.349 organize things alphabetically not at all it 00:02:18.349 --> 00:02:20.830 places data points based on their meaning so 00:02:20.830 --> 00:02:23.830 all the points about say fruits cluster together 00:02:23.830 --> 00:02:26.289 over here and animals form a totally different 00:02:26.289 --> 00:02:28.810 constellation over there when you ask what is 00:02:28.810 --> 00:02:31.550 a kitten your question itself gets turned into 00:02:31.550 --> 00:02:33.810 a new point of light. And it appears naturally 00:02:33.810 --> 00:02:36.710 right within that animal constellation. The database 00:02:36.710 --> 00:02:39.110 then super quickly finds the closest points, 00:02:39.189 --> 00:02:41.750 maybe cat, puppy, even wolf, and pulls the text 00:02:41.750 --> 00:02:44.189 link to them. That gives you a perfect, hyper 00:02:44.189 --> 00:02:45.909 -relevant answer. It's all about asking, what 00:02:45.909 --> 00:02:48.449 does this mean? And getting an intelligent, context 00:02:48.449 --> 00:02:50.969 -aware match. For a deep dive today, we'll be 00:02:50.969 --> 00:02:53.669 using Supabase as our vector database example, 00:02:53.889 --> 00:02:56.669 which is built right on top of Postgresful. So 00:02:56.669 --> 00:02:59.169 this database, it really understands concepts. 00:02:59.280 --> 00:03:02.219 then, not just keywords. Yes, precisely. It's 00:03:02.219 --> 00:03:04.680 all about meaning and context. Okay. And it's 00:03:04.680 --> 00:03:07.020 this clever organization that lets the two crucial 00:03:07.020 --> 00:03:10.159 halves of a RIAG system work so well together. 00:03:10.400 --> 00:03:13.139 If we kind of connect this to the bigger picture, 00:03:13.319 --> 00:03:16.050 building a RIAG system, it's like... assembling 00:03:16.050 --> 00:03:18.969 this, I don't know, magical, intelligent library, 00:03:19.110 --> 00:03:21.370 it's really a two -phase project, requires two 00:03:21.370 --> 00:03:23.889 distinct but totally complementary halves of 00:03:23.889 --> 00:03:25.650 a single brain, you could say. First, you've 00:03:25.650 --> 00:03:27.430 got what we call the AI librarian. That's your 00:03:27.430 --> 00:03:29.169 RA pipeline. This is the behind -the -scenes 00:03:29.169 --> 00:03:32.090 worker. Its job is to meticulously read every 00:03:32.090 --> 00:03:34.150 single document you give it, understand the content, 00:03:34.370 --> 00:03:36.409 and then place each piece of information on the 00:03:36.409 --> 00:03:39.449 correct shelf within your, well, infinitely large 00:03:39.800 --> 00:03:42.639 perfectly organized vector database. This isn't 00:03:42.639 --> 00:03:44.819 an ongoing job, mind you. It's a one -time process 00:03:44.819 --> 00:03:47.240 you run to initially stock the library. Right. 00:03:47.539 --> 00:03:50.319 And then you have the AI scholar. That's your 00:03:50.319 --> 00:03:52.860 R agent. This is the public -facing expert. The 00:03:52.860 --> 00:03:54.780 scholar kind of sits at the front desk, ready 00:03:54.780 --> 00:03:57.439 to help visitors. When you ask a difficult question, 00:03:57.620 --> 00:04:00.180 the scholar instantly zips through the library, 00:04:00.479 --> 00:04:02.599 pulls the exact pages from the correct books, 00:04:02.719 --> 00:04:05.080 synthesizes the information, and gives you this 00:04:05.080 --> 00:04:08.199 brilliant custom -written answer. But these two 00:04:08.199 --> 00:04:10.879 parts, they're completely interdependent. A scholar 00:04:10.879 --> 00:04:13.699 without a library is, well, just a generic chatbot. 00:04:13.800 --> 00:04:16.259 No specialized knowledge. And a library without 00:04:16.259 --> 00:04:18.680 a scholar, that's just a silent... inaccessible 00:04:18.680 --> 00:04:21.560 database. So one half is organizing all the knowledge 00:04:21.560 --> 00:04:23.439 and the other half leverages that organization 00:04:23.439 --> 00:04:26.420 to actually provide the answers. Precisely. A 00:04:26.420 --> 00:04:28.759 powerful two -phase system, yeah. So let's dig 00:04:28.759 --> 00:04:31.019 into the librarian's job then, that RAG pipeline. 00:04:31.240 --> 00:04:33.139 It's really all about transforming raw documents 00:04:33.139 --> 00:04:35.540 into a searchable, structured knowledge base. 00:04:35.759 --> 00:04:38.279 And this process has four key steps. First up, 00:04:38.379 --> 00:04:40.720 document input. This is simple. It's just acquiring 00:04:40.720 --> 00:04:42.819 the book. It's your source document. It would 00:04:42.819 --> 00:04:45.839 be a PDF, a text file, or even data you pull 00:04:45.839 --> 00:04:47.649 from other apps, maybe like a CRM. or something. 00:04:48.329 --> 00:04:50.490 Then we chunk the document. Think about it. You 00:04:50.490 --> 00:04:52.589 wouldn't file an entire 500 -page book under 00:04:52.589 --> 00:04:55.009 just one topic, right? You'd separate it into 00:04:55.009 --> 00:04:57.930 chapters or sections. Similarly, we split large 00:04:57.930 --> 00:05:01.149 documents into smaller, more focused pieces of 00:05:01.149 --> 00:05:03.050 text. This is super critical because it creates 00:05:03.050 --> 00:05:05.810 these context -rich pages for more accurate searching 00:05:05.810 --> 00:05:09.170 later on. A pro tip here, aim for chunks of maybe 00:05:09.170 --> 00:05:11.970 about 1 ,000 characters and use a 200 -character 00:05:11.970 --> 00:05:13.730 overlap between chunks. That just helps ensure 00:05:13.730 --> 00:05:15.629 concepts aren't awkwardly cut off right in the 00:05:15.629 --> 00:05:17.509 middle. Okay, and here's where it gets, well... 00:05:17.610 --> 00:05:21.110 really interesting. After chunking, you embed. 00:05:21.250 --> 00:05:23.750 This is I think the most magical part. An embeddings 00:05:23.750 --> 00:05:27.470 model. It's a sophisticated AI, acts like a universal 00:05:27.470 --> 00:05:30.430 translator. It converts those text chunks into 00:05:30.430 --> 00:05:32.550 vectors, which are basically just long strings 00:05:32.550 --> 00:05:35.269 of numbers. Think of it like giving each chunk 00:05:35.269 --> 00:05:38.230 a precise coordinate in that huge galaxy of meaning 00:05:38.230 --> 00:05:40.629 we talked about, that 1536 number you sometimes 00:05:40.629 --> 00:05:43.189 see. That just refers to the number of dimensions 00:05:43.189 --> 00:05:45.850 or kind of like traits an AI uses to describe 00:05:45.850 --> 00:05:47.990 each piece of information. It's like a super 00:05:47.990 --> 00:05:50.029 detailed profile for every single text chunk. 00:05:50.189 --> 00:05:52.910 And for your AI to understand your data consistently, 00:05:53.089 --> 00:05:56.180 this profile format, it has to match across all 00:05:56.180 --> 00:05:59.139 parts of your system. Finally, you vectorize 00:05:59.139 --> 00:06:01.839 those chunks. This just means storing these numerical 00:06:01.839 --> 00:06:04.639 vectors in your Supabase vector database. They 00:06:04.639 --> 00:06:07.060 get intelligently placed by meaning, right? So 00:06:07.060 --> 00:06:08.620 all the golf rules about putting are grouped 00:06:08.620 --> 00:06:10.980 together, far away from rules about driving. 00:06:11.439 --> 00:06:14.199 In a tool like N8AN, this whole process is actually 00:06:14.199 --> 00:06:16.660 simplified into like a five -node assembly line. 00:06:16.800 --> 00:06:18.699 You've got a trigger to start it, maybe a Google 00:06:18.699 --> 00:06:20.939 Drive node to get the document, a data loader 00:06:20.939 --> 00:06:23.500 for the chunking part, an embeddings model, like 00:06:23.500 --> 00:06:26.300 OpenAI's text embedding 3 small to do the translating, 00:06:26.439 --> 00:06:29.620 and the Supabase vector store node to file it 00:06:29.620 --> 00:06:31.519 all the way neatly. You just run this workflow 00:06:31.519 --> 00:06:33.939 once for each document you want to add. Gotcha. 00:06:34.000 --> 00:06:36.139 So it's about breaking down the info, translating 00:06:36.139 --> 00:06:39.300 it into this conceptual number language, and 00:06:39.300 --> 00:06:41.459 then filing it away smartly. Precisely. Creating 00:06:41.459 --> 00:06:43.860 that structured, searchable knowledge base. Yeah. 00:06:43.939 --> 00:06:46.019 Okay, to make this all super concrete, let's 00:06:46.019 --> 00:06:48.600 talk about our mission for today. We're building 00:06:48.600 --> 00:06:52.699 an AI golf caddy using a 22 -page PDF. It's called 00:06:52.699 --> 00:06:55.480 The Rules of Golf Simplified. That's our single 00:06:55.480 --> 00:06:58.060 source of truth. Why golf rules? Well, because 00:06:58.060 --> 00:07:00.879 they are incredibly dense, super specific, and 00:07:00.879 --> 00:07:04.399 totally self -contained. A generic AI. You might 00:07:04.399 --> 00:07:06.540 just guess or even hallucinate, you know, make 00:07:06.540 --> 00:07:09.139 stuff up if you ask a tricky golf question. But 00:07:09.139 --> 00:07:11.560 our engagement trained exclusively on this official 00:07:11.560 --> 00:07:14.860 rulebook. It'll be a true, reliable expert. Our 00:07:14.860 --> 00:07:17.339 goal is an AI caddy that can instantly and accurately 00:07:17.339 --> 00:07:19.720 answer very specific questions like, what am 00:07:19.720 --> 00:07:21.639 I allowed to do for practice? Or maybe, can I 00:07:21.639 --> 00:07:23.360 hit a practice shot between playing two holes? 00:07:23.600 --> 00:07:25.560 Yeah, and this isn't really about golf, is it? 00:07:25.740 --> 00:07:28.100 What's fascinating here is that this golf PDF, 00:07:28.420 --> 00:07:31.180 it's just a placeholder. This whole process is 00:07:31.180 --> 00:07:33.839 actually a universal blueprint for any specialized 00:07:33.839 --> 00:07:36.339 knowledge you want to give your AI. Your data 00:07:36.339 --> 00:07:39.839 source could be anything. Your last 5 ,000 HubSpot 00:07:39.839 --> 00:07:42.240 support tickets may be an air table base full 00:07:42.240 --> 00:07:44.519 of project data or even recent client emails. 00:07:45.279 --> 00:07:48.420 Imagine asking your AI, what are the top three 00:07:48.420 --> 00:07:50.660 common product issues we're seeing in the UK? 00:07:50.800 --> 00:07:53.660 Or, which projects are running over budget this 00:07:53.660 --> 00:07:56.259 year? and your agent's trigger, how you start 00:07:56.259 --> 00:07:58.379 the conversation. That can also be anything beyond 00:07:58.379 --> 00:08:00.680 just a chat window. It could be an email sent 00:08:00.680 --> 00:08:03.480 to askatyourcompany .com or a submission on your 00:08:03.480 --> 00:08:06.000 website form, or even a scheduled task that, 00:08:06.019 --> 00:08:08.120 say, summarizes the week's support tickets automatically. 00:08:08.600 --> 00:08:11.180 The skills you learn building this simple AI 00:08:11.180 --> 00:08:14.079 golf caddy, they're absolutely foundational for 00:08:14.079 --> 00:08:16.240 building really powerful business systems. So 00:08:16.240 --> 00:08:18.800 this simple golf example, it really unlocks vast 00:08:18.800 --> 00:08:21.230 possibilities for building custom AI then. Absolutely. 00:08:21.350 --> 00:08:23.329 It's a template, a template for pretty much any 00:08:23.329 --> 00:08:25.470 custom AI you can think of. All right, let's 00:08:25.470 --> 00:08:27.589 get our hands dirty then. Step one, building 00:08:27.589 --> 00:08:30.529 the library itself using Supabase. This is going 00:08:30.529 --> 00:08:32.870 to be the home for our AI agent. It's built on 00:08:32.870 --> 00:08:35.309 Postgresql, which is incredibly solid. Plus, 00:08:35.389 --> 00:08:37.909 it's pretty user -friendly and has a very generous 00:08:37.909 --> 00:08:40.070 free tier, which is great for getting started. 00:08:40.269 --> 00:08:43.529 Think of this step as laying the foundation, 00:08:43.830 --> 00:08:46.730 putting up the walls, and installing that magical 00:08:46.730 --> 00:08:48.889 shelving system for all your AI's knowledge. 00:08:49.419 --> 00:08:51.279 Okay, first up, you'll head over to supebase 00:08:51.279 --> 00:08:54.059 .com and create a new project. You give it a 00:08:54.059 --> 00:08:56.419 name, obviously, and critically, create a strong 00:08:56.419 --> 00:08:59.580 database password. Critical tip. Please, save 00:08:59.580 --> 00:09:01.659 this password securely right away, put it in 00:09:01.659 --> 00:09:03.419 a password manager, you will definitely need 00:09:03.419 --> 00:09:06.279 it later. Once your project is spinning up, might 00:09:06.279 --> 00:09:08.799 take a minute or two, you'll configure the vector 00:09:08.799 --> 00:09:11.590 database part. Now, this sounds a bit intimidating 00:09:11.590 --> 00:09:13.769 because it involves a snippet of code, but trust 00:09:13.769 --> 00:09:16.669 me, consider it a one -time magic spell. Exactly. 00:09:16.789 --> 00:09:18.710 A magic spell is a great way to put it. You just 00:09:18.710 --> 00:09:21.509 navigate to the SQL editor inside Superbase. 00:09:22.009 --> 00:09:24.470 You copy and paste a provided code block. We'll 00:09:24.470 --> 00:09:26.470 make sure you have access to that. And you click 00:09:26.470 --> 00:09:29.309 run that code. It's simply doing three things. 00:09:29.389 --> 00:09:31.950 First, create extension vector that's telling 00:09:31.950 --> 00:09:34.509 your database, hey, install the add -on to work 00:09:34.509 --> 00:09:36.370 with these AI vectors, these numbers representing 00:09:36.370 --> 00:09:40.409 meaning. Second, create table documents. This 00:09:40.409 --> 00:09:42.429 sets up the main storage shelf for your document 00:09:42.429 --> 00:09:44.909 chunks, including that special embedding vector 00:09:44.909 --> 00:09:47.950 1536 column we talked about earlier, the profile. 00:09:48.279 --> 00:09:51.360 Finally, create function match documents. This 00:09:51.360 --> 00:09:53.559 creates the magic search tool that finds similar 00:09:53.559 --> 00:09:56.200 documents by comparing their vectors, their meaning. 00:09:56.320 --> 00:09:59.000 That little bit that says one, documents .embedding, 00:09:59.019 --> 00:10:01.539 query embedding. That's just telling the database 00:10:01.539 --> 00:10:03.299 how to measure the conceptual distance between 00:10:03.299 --> 00:10:05.440 your question and every piece of info it has. 00:10:05.840 --> 00:10:08.120 Think of it like a reverse distance meter. The 00:10:08.120 --> 00:10:10.240 closer the concepts match, the higher the similarity 00:10:10.240 --> 00:10:12.580 score it gives. And then it brings back the highest 00:10:12.580 --> 00:10:14.779 scoring, most relevant information. So by running 00:10:14.779 --> 00:10:17.500 this one script, you instantly upgrade your plan. 00:10:17.580 --> 00:10:20.879 database into an ai powered knowledge base pretty 00:10:20.879 --> 00:10:23.480 cool right and the final part of this setup step 00:10:23.480 --> 00:10:26.159 is just getting the keys to the library so to 00:10:26.159 --> 00:10:28.740 speak in your super base project settings under 00:10:28.740 --> 00:10:31.320 the api section you'll find your project url 00:10:31.320 --> 00:10:33.980 and the service role secret key security warning 00:10:33.980 --> 00:10:37.059 this service role key it is the master key to 00:10:37.059 --> 00:10:39.519 your entire database treat it like a password 00:10:39.519 --> 00:10:43.190 like gold Never, ever share it publicly. You'll 00:10:43.190 --> 00:10:45.370 need both the URL and that key for connecting 00:10:45.370 --> 00:10:48.789 it in 8n later on. OK, so we're essentially building 00:10:48.789 --> 00:10:51.509 the AI's personal research library here, getting 00:10:51.509 --> 00:10:53.330 it all ready for use. That's it. Building the 00:10:53.330 --> 00:10:55.769 database and its intelligence shelves ready to 00:10:55.769 --> 00:10:59.149 be filled. Right. So with our library built and 00:10:59.149 --> 00:11:01.950 ready to be stocked or maybe already stocked, 00:11:01.950 --> 00:11:03.509 if we ran the pipeline, it's time to hire the 00:11:03.509 --> 00:11:06.690 scholar. Our ARIG agent. This is a new separate. 00:11:07.159 --> 00:11:09.879 and add in workflow. This one's for the interactive 00:11:09.879 --> 00:11:11.899 conversations. Think about the anatomy of this 00:11:11.899 --> 00:11:14.700 AI scholar. It's got the ears. That's a chat 00:11:14.700 --> 00:11:16.799 trigger node. It just listens for user questions. 00:11:17.139 --> 00:11:19.419 Then there's the central nervous system, an AI 00:11:19.419 --> 00:11:21.279 agent node that coordinates everything, the thinking 00:11:21.279 --> 00:11:23.759 and responding. The brain itself is an open AI 00:11:23.759 --> 00:11:26.940 chat model, maybe like GPT -40 mini, which does 00:11:26.940 --> 00:11:29.320 the actual thinking and generates the responses. 00:11:30.120 --> 00:11:32.120 Crucially, we give it the library card. That's 00:11:32.120 --> 00:11:34.879 a Supabase vector store tool. You add this as 00:11:34.879 --> 00:11:37.200 a tool specifically for the AI agent. You set 00:11:37.200 --> 00:11:39.740 its operation to retrieve documents. And you 00:11:39.740 --> 00:11:41.100 give it a really clear description, something 00:11:41.100 --> 00:11:43.059 like, use this tool to look up the official rules 00:11:43.059 --> 00:11:45.899 of golf. The AI needs that clarity. And finally, 00:11:45.980 --> 00:11:48.389 the universal translator. That's the embeddings 00:11:48.389 --> 00:11:50.509 model again. This must be the exact same model, 00:11:50.570 --> 00:11:53.309 say, text embedding three small that you use 00:11:53.309 --> 00:11:55.350 when you build a pipeline. That ensures consistent 00:11:55.350 --> 00:11:57.070 understanding for accurate searching. It's got 00:11:57.070 --> 00:11:59.129 to speak the same language. Now, a brilliant 00:11:59.129 --> 00:12:01.029 scholar who can't remember what you just said 00:12:01.029 --> 00:12:03.470 five seconds ago, that's incredibly frustrating, 00:12:03.690 --> 00:12:05.649 right? It's like the Dory from Finding Nemo problem. 00:12:06.110 --> 00:12:09.029 By default, your agent has zero short -term memory, 00:12:09.169 --> 00:12:11.389 so we need to give it a notepad. This is step 00:12:11.389 --> 00:12:14.750 four, adding conversational memory. And you can 00:12:14.750 --> 00:12:17.080 actually do this surprisingly easily. Right inside 00:12:17.080 --> 00:12:19.559 your AI agent node, you add the PostgreSQL chat 00:12:19.559 --> 00:12:22.379 memory feature. You'll create a new PostgreSQL 00:12:22.379 --> 00:12:24.779 credential using your Supabase database settings. 00:12:25.120 --> 00:12:27.960 Quick tip here, use the transaction pooler settings 00:12:27.960 --> 00:12:30.299 that Supabase provides. It helps with efficiency, 00:12:30.580 --> 00:12:32.620 especially if you plan to scale this up later. 00:12:33.529 --> 00:12:35.350 The password you need here is the one you set 00:12:35.350 --> 00:12:37.429 up for your Supabase project way back at the 00:12:37.429 --> 00:12:40.389 start. See? Told you you'd need it. You can also 00:12:40.389 --> 00:12:42.710 set the context window how many past messages 00:12:42.710 --> 00:12:45.570 it remembers. The default is five recent interactions, 00:12:45.710 --> 00:12:47.830 which is usually a good starting point. And this 00:12:47.830 --> 00:12:50.110 whole memory system works using session IDs. 00:12:50.389 --> 00:12:52.529 Think of it like a unique library card for each 00:12:52.529 --> 00:12:54.769 user. That ensures everyone gets their own separate 00:12:54.769 --> 00:12:57.009 continuous conversation. Your chat doesn't get 00:12:57.009 --> 00:13:00.370 mixed up with someone else's. I still wrestle 00:13:00.370 --> 00:13:03.389 with remembering to configure these. memory pieces 00:13:03.389 --> 00:13:05.269 myself sometimes it's just it's easy to miss 00:13:05.269 --> 00:13:07.870 yeah i can see that and testing this memory is 00:13:07.870 --> 00:13:10.090 pretty simple right you just start a new chat 00:13:10.090 --> 00:13:13.070 say hello my name is whatever your name is and 00:13:13.070 --> 00:13:16.409 then in the next message ask what's my name if 00:13:16.409 --> 00:13:20.029 it answers correctly bingo its notepad is working 00:13:20.029 --> 00:13:22.149 you can even double check by looking at the engine 00:13:22.149 --> 00:13:24.769 histories table directly in superbase if you 00:13:24.769 --> 00:13:27.389 want proof So this creates that conversational 00:13:27.389 --> 00:13:30.049 expert, one that can actually remember the previous 00:13:30.049 --> 00:13:33.289 turns in the conversation. Yes, exactly. It enables 00:13:33.289 --> 00:13:35.629 those interactive, context -aware conversations 00:13:35.629 --> 00:13:39.009 that feel much more natural. Okay, makes sense. 00:13:39.190 --> 00:13:41.350 Mid -roll sponsor read. Now let's talk about 00:13:41.350 --> 00:13:43.549 the fuel, the stuff that makes our AI engine 00:13:43.549 --> 00:13:47.919 actually run, the OpenAI API key. This is a step 00:13:47.919 --> 00:13:50.460 that can sometimes trip up beginners. It's really 00:13:50.460 --> 00:13:51.879 crucial to understand the difference between 00:13:51.879 --> 00:13:54.019 your ChatGPT Plus subscription, if you have one, 00:13:54.139 --> 00:13:57.480 and the OpenAI API. Think of ChatGP Plus like 00:13:57.480 --> 00:13:59.039 an all -you -can -eat buffet. You pay a flat 00:13:59.039 --> 00:14:02.179 fee, like $20 a month, for direct human -to -AI 00:14:02.179 --> 00:14:05.080 chatting through their website. The OpenAI API, 00:14:05.299 --> 00:14:07.679 though, that's different. It's a la carte. You 00:14:07.679 --> 00:14:10.080 pay for exactly what you use, usually charged 00:14:10.080 --> 00:14:12.000 per token, which is kind of like parts of words. 00:14:12.360 --> 00:14:15.120 This API is specifically designed for programmatic 00:14:15.120 --> 00:14:17.500 use, you know, machine to AI interaction. And 00:14:17.500 --> 00:14:19.620 that's exactly what you need for tools like NAI 00:14:19.620 --> 00:14:22.399 to work. Right. And getting your API key is pretty 00:14:22.399 --> 00:14:25.100 straightforward. You navigate to platform .opni 00:14:25.100 --> 00:14:27.779 .com. Notice the platform part. That's the developer 00:14:27.779 --> 00:14:30.799 side, not the regular chat GPT site. You'll create 00:14:30.799 --> 00:14:32.460 an account if you don't have one at a payment 00:14:32.460 --> 00:14:34.700 method. Think of it like opening a tab at a restaurant. 00:14:34.759 --> 00:14:36.740 You only pay for what you order. When you go 00:14:36.740 --> 00:14:40.019 to the API keys section, click create new secret 00:14:40.019 --> 00:14:42.120 key. It'll generate a key that starts with SCET. 00:14:42.259 --> 00:14:45.340 Security 101. Again, this API key is basically 00:14:45.340 --> 00:14:47.320 like a credit card number for your AI usage. 00:14:47.759 --> 00:14:50.179 Guard it fiercely. Copy it immediately. Save 00:14:50.179 --> 00:14:52.480 it in your secure password manager. Never, ever 00:14:52.480 --> 00:14:54.679 share it publicly. And once you close that little 00:14:54.679 --> 00:14:56.480 window where it shows you the key, you'll never 00:14:56.480 --> 00:14:58.580 see the full key again. If you lose it, you just 00:14:58.580 --> 00:15:01.000 have to generate a new one. Okay. And the golden 00:15:01.000 --> 00:15:03.919 rule of cost control here is really start small 00:15:03.919 --> 00:15:08.200 and set limits. Tip number one, use cheaper models 00:15:08.200 --> 00:15:10.860 when you're testing and building. GPT -4 -0 Mini 00:15:10.860 --> 00:15:13.220 is incredibly capable, especially now, and it's 00:15:13.220 --> 00:15:14.940 significantly cheaper than the bigger models 00:15:14.940 --> 00:15:17.500 like GPT -4 Turbo. You can always upgrade later 00:15:17.500 --> 00:15:19.360 once you know it works and you need more power. 00:15:19.940 --> 00:15:21.539 Tip number two, and this is probably the most 00:15:21.539 --> 00:15:24.820 important one, set hard usage limits. In your 00:15:24.820 --> 00:15:27.039 OpenAI billing settings, you can actually set 00:15:27.039 --> 00:15:29.639 a hard spending limit, maybe just $10 a month 00:15:29.639 --> 00:15:31.879 to start. If your usage ever hits that limit, 00:15:32.039 --> 00:15:34.139 the API will simply stop working until the next 00:15:34.139 --> 00:15:36.100 billing cycle. This is your ultimate safety net. 00:15:36.200 --> 00:15:38.000 It lets you experiment with total confidence, 00:15:38.179 --> 00:15:40.720 knowing you won't get a surprise bill. Got it. 00:15:40.840 --> 00:15:44.279 So the API key is the gateway to the AI's brainpower, 00:15:44.500 --> 00:15:47.000 and cost control is absolutely essential. Right. 00:15:47.179 --> 00:15:49.519 Access the brain, but keep that spending firmly 00:15:49.519 --> 00:15:52.220 in check. All right. The moment of truth, then. 00:15:52.360 --> 00:15:55.100 Let's test our AI golf caddy. You execute the 00:15:55.100 --> 00:15:57.379 agent workflow in NA8N. You open the chat interface 00:15:57.379 --> 00:15:59.419 it provides, and you ask a question. Let's try 00:15:59.419 --> 00:16:01.710 that one. What am I allowed to do for practice 00:16:01.710 --> 00:16:04.529 before a round? Now, behind the scenes, here's 00:16:04.529 --> 00:16:07.049 what happens. The agent gets your query. It identifies 00:16:07.049 --> 00:16:09.789 its look up the rules of golf tool as the right 00:16:09.789 --> 00:16:12.350 one to use. It sends your question to the embeddings 00:16:12.350 --> 00:16:14.669 model to get vectorized, turned into numbers. 00:16:14.850 --> 00:16:17.350 It searches this super -based database using 00:16:17.350 --> 00:16:19.590 those vector numbers. It finds the most relevant 00:16:19.590 --> 00:16:21.909 chunks of text about practice from that PDF we 00:16:21.909 --> 00:16:24.230 loaded. It feeds those relevant chunks to its 00:16:24.230 --> 00:16:26.929 GPD 4 .0 mini brain along with your original 00:16:26.929 --> 00:16:30.120 question. And then, it generates a perfect, context 00:16:30.120 --> 00:16:32.460 -aware answer based only on the provided rules. 00:16:32.759 --> 00:16:34.580 Yeah, and the response you'd likely get, pulled 00:16:34.580 --> 00:16:36.320 straight from the source document, would be pretty 00:16:36.320 --> 00:16:38.960 detailed. Something like, before a round, you 00:16:38.960 --> 00:16:40.820 can practice on the course on the day of a match 00:16:40.820 --> 00:16:43.000 play event, but not before a stroke play tournament 00:16:43.000 --> 00:16:46.080 or playoff, or between rounds, unless the committee 00:16:46.080 --> 00:16:48.980 allows it. During the round itself, no practice 00:16:48.980 --> 00:16:51.360 shots are allowed when playing a hole or between 00:16:51.360 --> 00:16:54.100 holes. Except for some chipping or putting on 00:16:54.100 --> 00:16:56.559 or near the last green you finished, a practice 00:16:56.559 --> 00:16:58.970 green, Or the next D -Box, provided it doesn't 00:16:58.970 --> 00:17:02.360 hold up play. Let's see. Super specific, right 00:17:02.360 --> 00:17:04.680 from the rules. And you can even check the agent's 00:17:04.680 --> 00:17:06.980 execution logs in ADN to see this whole thought 00:17:06.980 --> 00:17:09.460 process play out, which tool it used, what info 00:17:09.460 --> 00:17:12.079 it retrieved. It's great for debugging, but really 00:17:12.079 --> 00:17:14.380 this is just the hello world of RAG, you know. 00:17:14.539 --> 00:17:16.980 Beyond these basics, you can set up dynamic document 00:17:16.980 --> 00:17:19.740 updates. Imagine a workflow where any new file 00:17:19.740 --> 00:17:21.680 dropped in a Google Drive folder automatically 00:17:21.680 --> 00:17:23.960 triggers an update to the AI's knowledge base. 00:17:24.240 --> 00:17:26.700 For multi -user systems, you could use unique 00:17:26.700 --> 00:17:29.079 session IDs, like maybe a user's email address 00:17:29.079 --> 00:17:31.730 or phone number. to maintain potentially thousands 00:17:31.730 --> 00:17:33.750 of separate private conversations simultaneously. 00:17:34.849 --> 00:17:38.019 Beep. Whip. I mean, imagine scaling that. Giving 00:17:38.019 --> 00:17:41.160 a custom private AI brain to potentially millions 00:17:41.160 --> 00:17:43.400 of individual users, all running off one core 00:17:43.400 --> 00:17:46.799 system. That's truly remarkable when you think 00:17:46.799 --> 00:17:48.839 about it. And of course, there's always performance 00:17:48.839 --> 00:17:50.940 optimization you can do later. Things like fine 00:17:50.940 --> 00:17:53.220 -tuning the vector search parameters, maybe caching 00:17:53.220 --> 00:17:55.539 common queries to reduce API calls and costs. 00:17:55.799 --> 00:17:58.059 Lots you can do. So this example, it's really 00:17:58.059 --> 00:18:00.000 just scratching the surface of what's possible 00:18:00.000 --> 00:18:02.819 with this RAG approach. Absolutely. It's a powerful 00:18:02.819 --> 00:18:05.880 foundation. You can build almost infinite applications 00:18:05.880 --> 00:18:08.710 on top of this basic. So let's recap. You've 00:18:08.710 --> 00:18:11.369 just gone from a general chat bot, one that knows 00:18:11.369 --> 00:18:13.410 a lot about everything but nothing specific about 00:18:13.410 --> 00:18:16.670 your stuff, to an AI with a custom specialized 00:18:16.670 --> 00:18:20.910 brain. RDAG systems, powered by these clever 00:18:20.910 --> 00:18:24.289 vector databases, allow AI to do its own context 00:18:24.289 --> 00:18:26.829 -specific research within your data. You got 00:18:26.829 --> 00:18:29.779 the librarian. building the knowledge base, meticulously 00:18:29.779 --> 00:18:32.119 organizing the information, and the scholar using 00:18:32.119 --> 00:18:34.380 that knowledge base, retrieving precise facts 00:18:34.380 --> 00:18:36.779 to answer your questions accurately. And maybe 00:18:36.779 --> 00:18:38.960 the best part, you can actually build this all 00:18:38.960 --> 00:18:40.980 without writing a single line of traditional 00:18:40.980 --> 00:18:43.720 code, creating powerful, intelligent systems 00:18:43.720 --> 00:18:47.319 for anything from, well, golf rules to complex 00:18:47.319 --> 00:18:49.670 business data. Yeah, this isn't just some cool 00:18:49.670 --> 00:18:52.009 tech demo. It's really about building the future 00:18:52.009 --> 00:18:54.390 of how AI interacts with your unique world, your 00:18:54.390 --> 00:18:56.849 specific information. You've now got the foundational 00:18:56.849 --> 00:18:59.450 skills, the understanding to customize AI and 00:18:59.450 --> 00:19:01.549 build truly intelligent agents tailored to your 00:19:01.549 --> 00:19:04.109 needs. This deep dive has hopefully equipped 00:19:04.109 --> 00:19:06.329 you with a pretty profound understanding of how 00:19:06.329 --> 00:19:09.230 to give AI both a memory and a personal library. 00:19:09.849 --> 00:19:12.970 So the question for you is, what specific dense 00:19:12.970 --> 00:19:15.930 body of knowledge in your life or your work could 00:19:15.930 --> 00:19:18.529 you give your AI next? Imagine the expertise 00:19:18.529 --> 00:19:20.809 you could unlock. Thank you for diving deep with 00:19:20.809 --> 00:19:22.910 us today. Keep exploring, keep building, and 00:19:22.910 --> 00:19:23.869 yes, stay curious.