WEBVTT 00:00.300 --> 00:01.620 Eden: Hey there, Eden here. 00:01.620 --> 00:03.900 And in this video, we're going to discuss 00:03.900 --> 00:06.570 what's LangGraph and why do we need it, 00:06.570 --> 00:09.300 and why it's different from LangChain. 00:09.300 --> 00:12.150 So LangChain has been existing for a year 00:12.150 --> 00:14.040 and it's an amazing framework 00:14.040 --> 00:16.980 for building generative AI applications. 00:16.980 --> 00:20.670 If you want to build RAG applications or even agents, 00:20.670 --> 00:22.470 LangChain is here to help. 00:22.470 --> 00:25.260 It has gone a long way since it was created, 00:25.260 --> 00:26.850 and I mean it for the good. 00:26.850 --> 00:29.850 It got more secure, more flexible, 00:29.850 --> 00:32.100 more readable, and more usable, 00:32.100 --> 00:35.010 especially with the LangChain Expression Language. 00:35.010 --> 00:38.610 Leveraging composability, the way now to interact 00:38.610 --> 00:41.880 with the LangChain components couldn't be more convenient. 00:41.880 --> 00:45.180 And LangChain is the open-source framework 00:45.180 --> 00:47.670 which is at the moment pioneering 00:47.670 --> 00:50.160 a lot of work around agents. 00:50.160 --> 00:52.710 And in LangChain, we can build agents, 00:52.710 --> 00:55.050 and I even showed it in my LangChain course, 00:55.050 --> 00:58.860 how to build agents and what's the logic behind it. 00:58.860 --> 01:01.170 However, we do have limitations 01:01.170 --> 01:03.750 when we build agents within LangChain. 01:03.750 --> 01:05.370 And I think this diagram, 01:05.370 --> 01:08.730 which LangChain published when they launched LangGraph, 01:08.730 --> 01:10.680 demonstrating the challenges 01:10.680 --> 01:13.710 of building complex agentic systems. 01:13.710 --> 01:16.770 So, the person with the computer is code, 01:16.770 --> 01:21.150 and the person talking, you can think about it as an LLM. 01:21.150 --> 01:25.200 The top line represent sequential code that we write. 01:25.200 --> 01:29.310 This is deterministic code that simply runs and execute. 01:29.310 --> 01:33.060 So we, the developers, have full control of it. 01:33.060 --> 01:37.320 At the very bottom, we have agents or autonomous agents, 01:37.320 --> 01:41.430 which they can decide what they need to do and how to do it, 01:41.430 --> 01:42.870 and they simply do it. 01:42.870 --> 01:44.460 And by the way, at the moment, 01:44.460 --> 01:46.530 there aren't any autonomous agents 01:46.530 --> 01:49.593 which are quite production-ready or usable. 01:50.490 --> 01:52.920 So those have the most freedom. 01:52.920 --> 01:55.530 When we make a regular LLM call, 01:55.530 --> 01:58.860 then the LLM decides what is the output 01:58.860 --> 02:00.450 and we have no control of it, 02:00.450 --> 02:03.570 but we decide which steps to take before 02:03.570 --> 02:05.760 and which steps to take after. 02:05.760 --> 02:08.130 But those are very simple applications. 02:08.130 --> 02:11.310 For example, you can think of a cybersecurity company 02:11.310 --> 02:15.750 that is making a request to the LLM to explain an alert. 02:15.750 --> 02:18.930 We can build some useful features with it, 02:18.930 --> 02:21.630 but nothing really complex. 02:21.630 --> 02:24.030 And when we use LangChain chains, 02:24.030 --> 02:26.670 we can have a bunch of LLM calls 02:26.670 --> 02:29.520 and we still have control of what happens before 02:29.520 --> 02:30.960 and after those calls. 02:30.960 --> 02:33.750 So, we can still build some very cool things 02:33.750 --> 02:36.870 because we are chaining a lot of LLM calls together 02:36.870 --> 02:39.780 and we have more freedom at the moment. 02:39.780 --> 02:42.840 Now, within LangChain, we can use something 02:42.840 --> 02:46.110 which are called router chains or router agents, 02:46.110 --> 02:50.880 which are leveraging LLMs to decide which call to make 02:50.880 --> 02:52.950 or what steps to take. 02:52.950 --> 02:54.270 And we can easily do it 02:54.270 --> 02:55.830 with the LangChain Expression Language, 02:55.830 --> 02:57.960 and I showed it in my LangChain course. 02:57.960 --> 03:00.870 However, with the LangChain Expression Language, 03:00.870 --> 03:03.450 we can't make cycles. 03:03.450 --> 03:06.210 We can only have the acyclic graphs. 03:06.210 --> 03:09.810 So a certain flow that we write beforehand, 03:09.810 --> 03:12.480 but we can't iterate, we can't come back 03:12.480 --> 03:15.390 to the original node that we started from 03:15.390 --> 03:17.610 and start the process all over again. 03:17.610 --> 03:19.650 There are LangChain implementations 03:19.650 --> 03:20.880 for those kind of things, 03:20.880 --> 03:23.310 but they are ad hoc implementations, 03:23.310 --> 03:25.650 like for example, the React algorithm, 03:25.650 --> 03:29.160 where there is literally a while loop in the source code. 03:29.160 --> 03:32.160 This is exactly where LangGraph blends in. 03:32.160 --> 03:35.160 It gives us another dimension of freedom 03:35.160 --> 03:39.090 and complexity we can add into our agents. 03:39.090 --> 03:43.290 So we can with LangGraph implement cycles. 03:43.290 --> 03:45.090 And this feature is super important 03:45.090 --> 03:47.670 when we want to build very complex agent 03:47.670 --> 03:50.040 that have a certain amount of freedom 03:50.040 --> 03:52.560 that we are not used to from before. 03:52.560 --> 03:54.810 And the concept of this, 03:54.810 --> 03:57.840 it relates to something which is called flow engineering. 03:57.840 --> 04:00.120 So we as the developers, 04:00.120 --> 04:03.390 we can define the flow of our program 04:03.390 --> 04:06.293 and we can blend in with LangGraph the LLM. 04:06.293 --> 04:08.820 And the LLM can also help us 04:08.820 --> 04:10.740 where do we want to go in the flow, 04:10.740 --> 04:13.560 whether to go to flow A or flow B, 04:13.560 --> 04:17.250 or to finish, or to go back to where we started 04:17.250 --> 04:19.140 and to continue from there. 04:19.140 --> 04:22.230 So the cycles gives us a lot of freedom, 04:22.230 --> 04:25.500 and with LangGraph, it's very elegant 04:25.500 --> 04:28.713 and very easy to implement those kind of solutions. 04:29.610 --> 04:33.150 So from the LangGraph documentation, 04:33.150 --> 04:36.150 you can see it's listed as building language agents 04:36.150 --> 04:37.200 as graphs. 04:37.200 --> 04:41.160 So the entire logic, the entire flow of the agent 04:41.160 --> 04:44.250 can be expressed as a graph, 04:44.250 --> 04:46.080 as a graph with cycles, 04:46.080 --> 04:47.700 and this is very convenient. 04:47.700 --> 04:51.510 And with LangGraph, we can build some very advanced systems, 04:51.510 --> 04:54.483 and we'll even see something like that in the course.