WEBVTT 00:00.240 --> 00:01.680 Hey there Ethan here. 00:01.680 --> 00:06.720 And in this section we're going to build something which is called a reflection agent. 00:07.000 --> 00:13.480 It's going to extend our previous example of the reflection agent, but it's going to incorporate tools, 00:13.480 --> 00:19.280 for example, a search tool which is going to search online for real time data to enrich our answer. 00:19.640 --> 00:26.120 And we're going to review some advanced prompt engineering techniques in order for our reflection agent 00:26.160 --> 00:32.080 to be able to digest the feedback correctly and to really address and improve through the iterations, 00:32.320 --> 00:39.400 because to create a critique, it's not that hard, but to really leverage the LLM to incorporate that 00:39.400 --> 00:43.160 critique and to improve over time is something which is challenging. 00:43.160 --> 00:46.840 And for that, we're going to show you some very cool tricks that will help us do so. 00:47.240 --> 00:49.840 Now, where does this architecture come from? 00:49.840 --> 00:57.040 It comes from a paper which is called reflection, which is a joint paper from northeastern, MIT and 00:57.040 --> 00:57.720 Princeton. 00:58.000 --> 01:05.050 So the idea for this actually came from a lecture blog that their team created covering reflection agents 01:05.050 --> 01:07.450 and specifically covering the reflection paper. 01:07.490 --> 01:09.850 Implementing everything with Landgraaf. 01:10.490 --> 01:15.970 Now, I have to say that the link chain team and specifically Lance from Link Chain, they did an amazing 01:15.970 --> 01:16.410 work. 01:16.610 --> 01:23.570 However, their implementation is actually for me, it was a bit hard to understand, so it really took 01:23.570 --> 01:26.730 me a long time to understand what was going on. 01:26.930 --> 01:33.690 So what I did is take it and refactor it a bit in order to make it something which is much more easier 01:33.690 --> 01:39.490 to explain and easier to digest, because the implementation was quite hard to understand. 01:40.010 --> 01:43.850 And of course I'll be sharing the link in the video's resources section. 01:46.610 --> 01:53.290 The goal of the reflection agent is to give us a very detailed article about a topic that will give 01:53.290 --> 01:58.930 it, and we want the article to dynamically fetch relevant information from the web. 01:59.210 --> 02:02.650 And we want to have citations for the reference data. 02:02.650 --> 02:06.610 And of course we want to incorporate a quality right critique loop. 02:07.020 --> 02:09.620 We basically want to get a very high quality answer. 02:10.900 --> 02:18.060 And this is an example question right, about AI powered SoC, autonomous SoC problem domain startups 02:18.100 --> 02:20.420 that do that and raised capital. 02:20.980 --> 02:28.020 So I'm not sure if you're familiar, but AI powered SoC or Hyperautomation autonomous SoC like companies 02:28.020 --> 02:34.660 like talk is something which is exploding right now, and it's getting a lot of attention. 02:35.020 --> 02:41.860 And the idea here is to take the security operations center and to leverage AI agents in order to start 02:41.860 --> 02:46.260 resolving tier one tickets and security incidents. 02:46.260 --> 02:52.420 So those kind of tickets don't require a lot of reasoning, and they can leverage external tools in 02:52.420 --> 02:59.260 order to triage them and to resolve them, and freeing up a lot of work for SOC analysts already. 02:59.260 --> 03:02.900 This is an example response we got from the reflection agent. 03:03.140 --> 03:07.060 And we're going now to discuss the architecture of that agent. 03:08.540 --> 03:14.550 So the architecture looks very similar to our reflection agent we saw in the previous section. 03:15.110 --> 03:22.630 You can see when we start the agent flow, then we first have a responder node which is going to respond. 03:22.670 --> 03:23.750 The initial response. 03:23.790 --> 03:25.110 You can see it right over here. 03:25.710 --> 03:33.070 However it's not only generates the original response, but it also adds a critique to the response 03:33.070 --> 03:35.950 itself and a search term. 03:36.270 --> 03:43.510 So in the search term, the agent is going to come up with ideal search queries that would be beneficial 03:43.510 --> 03:45.150 to get a better response. 03:45.910 --> 03:53.150 It's going to help us ground the output that we gave with current events and external data that is available 03:53.190 --> 03:53.750 online. 03:54.430 --> 04:00.030 So after we have the search queries, we're going to execute the execute tools node, which is going 04:00.030 --> 04:01.310 to take our search queries. 04:01.310 --> 04:05.990 And simply going to use a search engine to retrieve us results in real time. 04:06.310 --> 04:10.710 So I'm going to be using Tavileh, which is an amazing third party. 04:10.750 --> 04:18.130 That is a search engine which is highly optimized for LM applications, and we can easily downstream 04:18.130 --> 04:20.930 the responses that we get to the LM. 04:21.290 --> 04:27.530 And now we're going to downstream the results, and we're going to go to the Revisor node. 04:27.530 --> 04:33.570 And the Revisor node is going to take the initial response, which has already an initial critique, 04:33.890 --> 04:37.290 and the results of the search engine execution. 04:37.290 --> 04:39.130 So it has also external data. 04:39.330 --> 04:42.330 It is very relevant to the topic that we're executing. 04:42.730 --> 04:46.370 It's going to revisit and change our original response. 04:46.530 --> 04:51.450 Now it's going to do that while incorporating the new data and the critique. 04:51.450 --> 04:57.650 So it's going to address the suggestion that we got in the previous step to articulate a better answer. 04:57.650 --> 05:04.370 But not only that, the Revisor is also going to supply us a new critique to the new revision of the 05:04.370 --> 05:11.050 article, and it's going to provide us new search terms that we want to now look up, according to the 05:11.050 --> 05:17.940 revised articles that are going to be beneficial, and it's going to give us The citations of the first 05:17.940 --> 05:19.060 search that we had. 05:20.100 --> 05:24.580 And after that we're going to continue and search for the new queries. 05:24.740 --> 05:30.900 We're going to downstream the new information alongside with the critique, and we're going to revise 05:30.900 --> 05:31.580 it again. 05:31.580 --> 05:36.300 So this loop is going to keep happening until we hit a stopping condition. 05:36.580 --> 05:38.460 And after that we finished. 05:38.660 --> 05:44.540 And this architecture is very similar to our reflection agent architecture that we saw in the previous 05:44.540 --> 05:45.180 section. 05:45.620 --> 05:47.700 But we added here a search engine. 05:49.060 --> 05:50.820 Let's discuss what we're going to use. 05:50.820 --> 05:57.500 We're going to use GPT four turbo because we need a strong enough model to write the text and to write 05:57.500 --> 06:01.380 the critique, which have a good reasoning power. 06:02.060 --> 06:07.220 Also be leveraging function calling, which is going to be super important in this implementation. 06:07.660 --> 06:10.820 And we're going to be using Tavileh as our search engine. 06:10.820 --> 06:17.580 And of course, we're going to be using Lindsmith for tracing, because in this complex architecture 06:17.700 --> 06:20.180 we want to be able to trace easily.