WEBVTT 00:00.880 --> 00:01.520 Hi everyone. 00:01.520 --> 00:02.400 Welcome back. 00:03.120 --> 00:10.400 Today we are diving into RAC which stands for Retrieval Augmented Generation and is a powerful technique 00:10.400 --> 00:13.360 that makes AI agents and automation way smarter. 00:14.240 --> 00:20.720 By the end of this lesson, you will know what the RAC is and why it's used so useful, how it retrieves 00:20.720 --> 00:26.120 and generates information dynamically, and how to build a RAC powered workflow in an A-10. 00:26.880 --> 00:28.120 So let's jump right in. 00:32.200 --> 00:35.160 First, what is retrieval augmented generation? 00:36.400 --> 00:37.560 Think of it like this. 00:38.240 --> 00:42.120 Large language models only use what they have been trained on. 00:42.560 --> 00:49.200 So if you ask them about something new, they might just guess or worse, hallucinate. 00:50.280 --> 00:54.920 So RAC fixes this by adding a retrieval step before generating a response. 00:56.240 --> 01:03.730 So retrieval means it fetches the most relevant up to date info from databases, documents, sheets, 01:04.330 --> 01:05.650 APIs, etc.. 01:07.010 --> 01:13.810 Generation means the AI then uses that info to create a better, more accurate response. 01:14.570 --> 01:22.290 For example, imagine you ask a powered assistant about today's stock prices. 01:22.930 --> 01:32.570 A normal chatbot like ChatGPT might give you outdated info even if you are using ChatGPT search. 01:33.010 --> 01:35.610 It might still make mistakes. 01:36.770 --> 01:40.050 That's why it's better to use perplexity for searching the web. 01:44.050 --> 01:45.730 Now why is that important? 01:47.130 --> 01:51.170 So most llms have a fixed knowledge base. 01:52.210 --> 02:00.070 They don't know anything after the last training update, even if they have access to tools and the 02:00.070 --> 02:00.750 internet. 02:01.270 --> 02:04.790 They don't have access to your internal private documents, right? 02:05.470 --> 02:13.910 So if rack, we can use external sources of data instead of relying on outdated training data, you 02:13.910 --> 02:21.910 can pull in real time information like stock prices, news or customer support data. 02:23.230 --> 02:27.790 And rack makes AI more accurate by reducing hallucinations. 02:29.230 --> 02:29.670 All right. 02:29.670 --> 02:33.030 So now let's break down how rack works in simple terms. 02:33.390 --> 02:37.270 Imagine you ask your rack powered AI assistant a question. 02:37.870 --> 02:42.510 Let's say what what's the latest refund policy for my company? 02:44.310 --> 02:45.350 First the LLM. 02:46.510 --> 02:54.030 So your AI assistant brain analyzes your question and figures out what information it needs. 02:54.630 --> 03:02.560 Then instead of making an answer straight away, it searches through connected data sources, and this 03:02.560 --> 03:07.000 could be internal documentation, a database or a knowledge base. 03:09.160 --> 03:14.560 Then the system retrieves the most relevant information related to your question. 03:15.520 --> 03:23.240 And then the LLM combines that information with its own language understanding to generate a factual 03:23.280 --> 03:25.200 and relevant response. 03:26.800 --> 03:33.760 And finally, the AI gives you an answer that's backed by real data and in some cases even includes 03:33.760 --> 03:37.280 references to the sources it used. 03:38.440 --> 03:45.880 For example, instead of saying I'm not sure, an AI assistant followed by RAC for customer support 03:46.160 --> 03:54.040 retrieves the exact refund policy from company documents and provides an accurate response to the customer. 03:54.440 --> 03:57.960 It will be much easier to understand by looking at this diagram. 03:57.960 --> 04:10.250 Um, so let's say a customer asks chatbot, what's the warranty policy for my new laptop and assistant, 04:10.250 --> 04:14.130 powered by rack handles it step by step. 04:18.850 --> 04:22.370 So the customer types the question into the chat. 04:23.530 --> 04:29.490 Then the chatbot processes the request and determines it needs warranty details. 04:29.690 --> 04:36.930 So instead of relying on old pre-trained responses, it performs a real time search. 04:37.770 --> 04:48.650 So it retrieves the latest warranty policy from the company's internal database and if needed and if 04:48.650 --> 04:53.450 needed, it checks the company's online help center for updates. 04:53.930 --> 05:00.180 It can even search the web for manufacturer Enhancements or regulation changes. 05:00.620 --> 05:04.980 So the assistant gathers the most relevant data. 05:05.500 --> 05:09.740 And then generates an accurate response based on the latest information. 05:10.220 --> 05:16.980 So the customer instantly gets the most up to date answer without needing to call support or search 05:17.020 --> 05:17.740 manually. 05:19.020 --> 05:26.100 It's really helpful because if the company updated its return policy yesterday, the AI assistant powered 05:26.100 --> 05:30.300 by RAC will instantly reflect the changes. 05:31.140 --> 05:33.980 Now, how does RAC work in your company? 05:34.540 --> 05:42.020 For example, imagine you build an AI assistant for your company's customer support and sales team. 05:42.500 --> 05:43.820 And is powered by RAC. 05:45.460 --> 05:50.100 And in this scenario, a customer asks what are your refund policies? 05:50.540 --> 05:53.900 Instead of answering I don't know, like ChatGPT might. 05:53.940 --> 06:03.640 If it lacks the data and it's not fine tuned, or it can hallucinate and give you a random response. 06:06.360 --> 06:14.760 Instead of that, your AI assistant powered by RAC retrieves the latest refund policy from your company's 06:14.800 --> 06:17.280 knowledge base or internal documentation. 06:17.600 --> 06:27.480 So it summarizes the key points like refund eligibility, like required steps, etc. and the customer 06:27.520 --> 06:30.320 gets precise and up to date answer. 06:31.000 --> 06:33.240 And it really reduces confusion. 06:34.200 --> 06:40.040 Now let's jump into N810 and explore the components of RAC Powered assistant. 06:41.720 --> 06:46.240 You can build this RAC assistant relatively easily in N810. 06:46.440 --> 06:48.880 And I will show you exactly how in the next lesson. 06:49.880 --> 06:53.200 It might look complex, but trust me, it's super simple. 06:53.920 --> 06:58.490 And this particular example is very practical and very useful. 07:00.210 --> 07:04.250 So here is how it works step by step. 07:06.170 --> 07:09.170 Assistant which is our tools agent. 07:09.370 --> 07:10.970 This is the core of the workflow. 07:11.250 --> 07:20.530 So inside it has a system prompt which is a set of instructions that tell our assistant what to do. 07:21.010 --> 07:24.290 It decides how to process a user queries. 07:24.290 --> 07:28.570 So basically what we expect from this assistant to do. 07:29.690 --> 07:31.770 Next we have chat model. 07:31.770 --> 07:33.210 So OpenAI chat model. 07:33.490 --> 07:36.490 This node is the brain of our assistant. 07:37.250 --> 07:39.170 Inside this node we can select. 07:41.210 --> 07:42.730 Between different llms. 07:43.090 --> 07:48.050 And this is the AI that actually generates responses. 07:49.770 --> 07:54.650 It doesn't just guess it waits until it has the right information before answering. 07:55.690 --> 08:00.180 Not based on the ChatGPT general knowledge. 08:03.020 --> 08:05.100 Now Postgres chat memory. 08:05.100 --> 08:09.300 So this acts as our assistance persistent memory. 08:09.300 --> 08:15.100 So it helps our assistant remember past interactions with users, which makes conversations feel more 08:15.100 --> 08:18.660 natural because it knows the context of the conversation. 08:20.420 --> 08:22.900 Now we have Supabase vector store. 08:22.940 --> 08:25.940 This is where all the company knowledge is stored. 08:26.220 --> 08:29.660 Basically, this is part of a vector database. 08:30.500 --> 08:32.740 We'll dive into vector databases later. 08:33.300 --> 08:43.940 For now, just know that it acts like an AI powered search engine and helps our LLM retrieve the most 08:43.940 --> 08:47.940 relevant information before the chatbot responds. 08:50.020 --> 08:51.140 Now embeddings. 08:51.940 --> 09:00.480 This node is basically the setup which converts text into a format that the AI model. 09:00.720 --> 09:04.040 So large language model can search through quickly. 09:04.360 --> 09:08.200 It turns human language into numerical representations. 09:08.720 --> 09:16.800 So when a user asks a question llms search through embeddings to find the most relevant data in the 09:16.800 --> 09:17.600 vector store. 09:18.760 --> 09:21.920 So we'll explore embeddings in more, more detail later. 09:22.240 --> 09:24.760 But for now, just know that they help. 09:24.800 --> 09:28.960 I understand and compare text efficiently. 09:29.880 --> 09:34.440 Now let's talk about how react keeps AI models. 09:34.480 --> 09:36.920 So large language models are useful. 09:37.840 --> 09:45.800 So most AI assistants either guess answers or rely on old pre-trained data. 09:46.760 --> 09:50.040 That's a problem if you need real time accurate info. 09:50.480 --> 09:55.120 So react solves this by pulling data from the right place at the right time. 09:57.050 --> 10:00.290 Basically it search internal documents. 10:00.610 --> 10:04.770 Let's say an employee asks how many sick days do I get? 10:05.050 --> 10:14.570 So instead of just making some random response, our assistant checks the official HR handbook and delivers 10:14.570 --> 10:16.850 the exact policy details. 10:19.170 --> 10:22.210 Our RAC powered assistant can retrieve web data. 10:22.210 --> 10:28.770 So, for example, if a customer asks what's the latest price for this laptop? 10:29.890 --> 10:31.450 Our assistant doesn't guess. 10:31.450 --> 10:36.770 It pulls live pricing from the store's website and provides the real answer. 10:38.010 --> 10:44.370 Now, it also adapts to fast changing industries like a finance assistant. 10:45.650 --> 10:51.650 Let's say it gets asked what's the new tax rule for freelancers? 10:51.970 --> 11:00.620 So instead of giving outdated info, It retrieves the latest IRS update and provides an accurate response. 11:01.220 --> 11:08.660 So in most cases, instead of guessing or saying I don't know, it fetches the answer from the right 11:08.700 --> 11:10.940 source, which is really useful. 11:11.780 --> 11:17.980 So now these are the most popular use cases for using crack. 11:19.060 --> 11:21.300 Where does crack actually make a difference? 11:22.540 --> 11:30.180 First of all, in customer support, because instead of a chatbot giving a generic check our website 11:30.180 --> 11:33.780 response, it fetches real answers from past tickets. 11:34.380 --> 11:35.740 Frequently asked questions. 11:36.580 --> 11:37.900 Past conversations. 11:38.700 --> 11:40.300 Customer history, etc.. 11:41.620 --> 11:45.220 It is also very useful in chatbots and virtual assistants. 11:46.020 --> 11:50.500 Think of a virtual assistant that doesn't just repeat old info. 11:50.540 --> 11:57.990 It grabs live updates on news, weather or company policies before responding. 12:01.070 --> 12:03.350 It's often used in education and research. 12:04.510 --> 12:12.670 So for students looking for a concept or a lawyer searching for the latest case law, react can retrieve 12:12.670 --> 12:15.590 the most relevant documents very quickly. 12:15.950 --> 12:22.670 It is also very useful in e-commerce, so it can check livestock availability and even suggest personalized 12:22.670 --> 12:25.270 recommendations based on past orders. 12:29.910 --> 12:35.150 It's well known in finance and banking, so customers don't need to dig through statements. 12:35.190 --> 12:41.990 I can instantly retrieve account balances, low options or interest rates in real time. 12:43.470 --> 12:50.150 Next, healthcare and insurance so it can pull up your actual policy info before answering. 12:51.670 --> 12:53.890 And lastly, legal and compliance. 12:54.810 --> 13:02.530 So instead of providing outdated info, it can grab the latest regulations and contract updates before 13:02.890 --> 13:04.410 giving an advice. 13:04.930 --> 13:08.090 All right, so let's wrap up what we have covered today. 13:09.050 --> 13:18.410 RAC retrieves real data before generating answers, which makes Nlm's responses way more accurate. 13:18.890 --> 13:20.490 It can pull from internal. 13:21.210 --> 13:29.490 It can pull data from internal documents, websites and databases, keeping responses more accurate 13:29.730 --> 13:30.610 and up to date. 13:32.610 --> 13:40.890 Assistance powered by RAC can give you real, useful information, and you can build RAC powered assistance 13:40.890 --> 13:43.690 yourself using N810 relatively easily. 13:45.130 --> 13:47.770 The next lesson we'll dive into N810. 13:48.890 --> 13:51.090 So stick around and let's start cooking. 13:52.090 --> 13:53.010 See you in the next one!