1 00:00:07,000 --> 00:00:11,000 So we all know what Retrieval Augmented Generation is. 2 00:00:11,000 --> 00:00:13,000 Let's just do a quick refresher. 3 00:00:13,000 --> 00:00:17,000 Retrieval Augmented Generation is a powerful and popular pipeline 4 00:00:17,000 --> 00:00:20,000 that enhances responses from a large language model. 5 00:00:20,000 --> 00:00:24,000 It does this by incorporating relevant data retrieved from a vector database, 6 00:00:24,000 --> 00:00:29,000 adding it as context to the prompt, and sending it to the LLM for generation. 7 00:00:29,000 --> 00:00:34,000 What this does is it allows the LLM to ground its response in concrete and accurate information, 8 00:00:34,000 --> 00:00:38,000 and that improves the quality and reliability of the response. 9 00:00:38,000 --> 00:00:40,000 Let me quickly sketch it out. 10 00:00:40,000 --> 00:00:47,000 So let's say we have a user, or an application even, 11 00:00:47,000 --> 00:00:51,000 and they send a query. 12 00:00:51,000 --> 00:00:53,000 Now, without Retrieval Augmented Generation, 13 00:00:53,000 --> 00:01:01,000 this query is going to go and get itself interpolated into a prompt. 14 00:01:01,000 --> 00:01:07,000 And from there, that's going to hit the LLM, 15 00:01:07,000 --> 00:01:14,000 and that's going to generate an output. 16 00:01:14,000 --> 00:01:18,000 To make this RAG, we can add a vector database. 17 00:01:18,000 --> 00:01:21,000 So instead of just going directly and getting itself interpolated into the prompt, 18 00:01:21,000 --> 00:01:24,000 it's going to hit this vector DB, 19 00:01:24,000 --> 00:01:29,000 and the response from that vector DB is going to be used as context for the prompt. 20 00:01:29,000 --> 00:01:33,000 Now, in this typical RAG pipeline, we call the LLM only once, 21 00:01:33,000 --> 00:01:37,000 and we use it solely to generate a response. 22 00:01:37,000 --> 00:01:41,000 But what if we could leverage the LLM not just for responses, 23 00:01:41,000 --> 00:01:43,000 but also for additional tasks, 24 00:01:43,000 --> 00:01:47,000 like deciding which vector database to query if we have multiple databases, 25 00:01:47,000 --> 00:01:50,000 or even determining the type of response to give? 26 00:01:50,000 --> 00:01:54,000 Should it answer with text, generate a chart, or even provide a code snippet? 27 00:01:54,000 --> 00:01:58,000 And that would all be dependent on the context of that query. 28 00:01:58,000 --> 00:02:05,000 So this is where the agentic RAG pipeline comes into play. 29 00:02:05,000 --> 00:02:09,000 In agentic RAG, we use the LLM as an agent, 30 00:02:09,000 --> 00:02:12,000 and the LLM goes beyond just generating a response. 31 00:02:12,000 --> 00:02:16,000 It takes on an active role and can make decisions that will improve both 32 00:02:16,000 --> 00:02:19,000 the relevance and accuracy of the retrieved data. 33 00:02:19,000 --> 00:02:24,000 Now let's explore how we can augment the initial process with an agent 34 00:02:24,000 --> 00:02:26,000 and a couple of different sources of data. 35 00:02:26,000 --> 00:02:32,000 So instead of just one single source, let's add a second. 36 00:02:32,000 --> 00:02:38,000 And the first one can be, you know, internal documentation, right? 37 00:02:38,000 --> 00:02:46,000 And the second one can be general industry knowledge. 38 00:02:46,000 --> 00:02:49,000 Now in the internal documentation, we're going to have things like policies, 39 00:02:49,000 --> 00:02:50,000 procedures, and guidelines. 40 00:02:50,000 --> 00:02:54,000 And the general knowledge base will have things like industry standards, 41 00:02:54,000 --> 00:02:57,000 best practices, and public resources. 42 00:02:57,000 --> 00:03:01,000 So how can we get the LLM to use the vector database that contains the data 43 00:03:01,000 --> 00:03:04,000 that will be most relevant to the query? 44 00:03:04,000 --> 00:03:11,000 Let's add that agent into this pipeline. 45 00:03:11,000 --> 00:03:15,000 Now this agent can intelligently decide which database to query 46 00:03:15,000 --> 00:03:17,000 based on the user's question. 47 00:03:17,000 --> 00:03:19,000 And the agent isn't making a random guess. 48 00:03:19,000 --> 00:03:23,000 It's leveraging the LLM's language understanding capabilities 49 00:03:23,000 --> 00:03:27,000 to interpret the query and determine its context. 50 00:03:27,000 --> 00:03:29,000 So if an employee asks, 51 00:03:29,000 --> 00:03:32,000 "What's the company's policy on remote work during the holidays?" 52 00:03:32,000 --> 00:03:34,000 It would route that to the internal documentation. 53 00:03:34,000 --> 00:03:38,000 And that response will be used as context for the prompt. 54 00:03:38,000 --> 00:03:40,000 But if the question is more general, like, 55 00:03:40,000 --> 00:03:44,000 "What are the industry standards for remote work in tech companies?" 56 00:03:44,000 --> 00:03:47,000 The agent's going to route that to the general knowledge database. 57 00:03:47,000 --> 00:03:50,000 And that context is going to be used within that prompt. 58 00:03:50,000 --> 00:03:53,000 Powered by an LLM and properly trained, 59 00:03:53,000 --> 00:03:55,000 the agent analyzes the query, 60 00:03:55,000 --> 00:03:58,000 and based on the understanding of the content and the context, 61 00:03:58,000 --> 00:04:00,000 decides which database to use. 62 00:04:00,000 --> 00:04:03,000 But they're not always going to ask questions that are generally 63 00:04:03,000 --> 00:04:07,000 or genuinely relevant to any of the stuff that we have in our vector DB. 64 00:04:07,000 --> 00:04:11,000 So what if someone asks a question that is just totally out of left field? 65 00:04:11,000 --> 00:04:14,000 Like, "Who won the World Series in 2015?" 66 00:04:14,000 --> 00:04:19,000 What the agent can do at that point is it could route it to a fail-safe. 67 00:04:21,000 --> 00:04:27,000 So because the agent is able to recognize the context of the query, 68 00:04:27,000 --> 00:04:31,000 it could recognize that it's not a part of the two databases that we have. 69 00:04:31,000 --> 00:04:34,000 It could route it to the fail-safe and return back, 70 00:04:34,000 --> 00:04:38,000 "Sorry, I don't have the information you're looking for." 71 00:04:38,000 --> 00:04:43,000 This agentic RAG pipeline can be used in customer support systems and legal tech. 72 00:04:43,000 --> 00:04:46,000 For example, a lawyer can source answers to their questions 73 00:04:46,000 --> 00:04:48,000 from their internal briefs, 74 00:04:48,000 --> 00:04:52,000 and then in another query, just get stuff from public caseload databases. 75 00:04:52,000 --> 00:04:55,000 The agent can be utilized in a ton of ways. 76 00:04:55,000 --> 00:04:59,000 Agentic RAG is an evolution in how we enhance the RAG pipeline 77 00:04:59,000 --> 00:05:02,000 by moving beyond simple response generation 78 00:05:02,000 --> 00:05:04,000 to more intelligent decision-making. 79 00:05:04,000 --> 00:05:07,000 By allowing an agent to choose the best data sources 80 00:05:07,000 --> 00:05:10,000 and potentially even incorporate external information, 81 00:05:10,000 --> 00:05:13,000 like real-time data or third-party services, 82 00:05:13,000 --> 00:05:16,000 we can create a pipeline that's more responsive, 83 00:05:16,000 --> 00:05:19,000 more accurate, and more adaptable. 84 00:05:19,000 --> 00:05:21,000 This approach opens up so many possibilities 85 00:05:21,000 --> 00:05:25,000 for applications in customer service, legal tech, healthcare, 86 00:05:25,000 --> 00:05:27,000 virtually any field. 87 00:05:27,000 --> 00:05:29,000 As the technology continues to evolve, 88 00:05:29,000 --> 00:05:32,000 we will see AI systems that truly understand context 89 00:05:32,000 --> 00:05:34,000 and can deliver amazing values.