WEBVTT 1 00:00:00.330 --> 00:00:01.440 Hey there, Eden here, 2 00:00:01.440 --> 00:00:04.980 and I want to have a recap on what we did 3 00:00:04.980 --> 00:00:08.520 so far in the course when it comes to agents. 4 00:00:08.520 --> 00:00:13.520 And by now, I really hope that the ReAct algorithm 5 00:00:13.860 --> 00:00:16.470 architecture and implementation 6 00:00:16.470 --> 00:00:19.830 and everything around it is crystal clear. 7 00:00:19.830 --> 00:00:22.950 And I'm hoping that if I'm going to wake you up at night 8 00:00:22.950 --> 00:00:26.280 and ask you, "What's the flow of the ReAct algorithm?" 9 00:00:26.280 --> 00:00:28.650 then you'll be able to sing it. 10 00:00:28.650 --> 00:00:30.780 And I'll have a really quick recap. 11 00:00:30.780 --> 00:00:34.230 We send the query to the ReAct agent, 12 00:00:34.230 --> 00:00:36.930 the large language model then ponders 13 00:00:36.930 --> 00:00:40.260 and thinks and decides which tool to call. 14 00:00:40.260 --> 00:00:42.630 Then we go and execute that tool 15 00:00:42.630 --> 00:00:46.860 and then we go and do that until we get a final answer 16 00:00:46.860 --> 00:00:49.710 until there are no more tools to invoke. 17 00:00:49.710 --> 00:00:53.250 All right, so let's go back to the diagram 18 00:00:53.250 --> 00:00:55.230 that I showed you earlier in the course 19 00:00:55.230 --> 00:00:58.830 about the evolution of ReAct agents. 20 00:00:58.830 --> 00:01:03.030 So it all started with the ReAct paper 21 00:01:03.030 --> 00:01:05.220 and the ReAct prompt, 22 00:01:05.220 --> 00:01:09.480 where we used the LLM as a reasoning engine 23 00:01:09.480 --> 00:01:14.130 and then LangChain implemented some fancy parsing 24 00:01:14.130 --> 00:01:18.330 in order to extract which tools to execute. 25 00:01:18.330 --> 00:01:20.640 Okay, so this was the very beginning. 26 00:01:20.640 --> 00:01:23.460 And this implementation was super impressive, 27 00:01:23.460 --> 00:01:25.680 but it wasn't that reliable 28 00:01:25.680 --> 00:01:28.230 so we couldn't really use this in production 29 00:01:28.230 --> 00:01:32.460 because the models weren't that strong by that time 30 00:01:32.460 --> 00:01:36.060 and it was really hard to parse the output. 31 00:01:36.060 --> 00:01:38.700 Because the models weren't that reliable, 32 00:01:38.700 --> 00:01:41.430 then parsing the output was really hard. 33 00:01:41.430 --> 00:01:43.410 It was really non-deterministic 34 00:01:43.410 --> 00:01:45.450 so we really didn't have control 35 00:01:45.450 --> 00:01:48.180 about which token the LLM is going to generate. 36 00:01:48.180 --> 00:01:50.820 And because of that, it's enough that the LLM 37 00:01:50.820 --> 00:01:53.490 is going to generate one wrong token, 38 00:01:53.490 --> 00:01:57.660 and it can cause and mess up all the output parsing. 39 00:01:57.660 --> 00:02:01.290 And then LLMs became a bit better 40 00:02:01.290 --> 00:02:03.120 and function calling came out, 41 00:02:03.120 --> 00:02:06.960 and function calling normalized this process 42 00:02:06.960 --> 00:02:11.190 of making the LLM behave as a reasoning engine 43 00:02:11.190 --> 00:02:15.240 and suddenly we don't need this ReAct prompt anymore, 44 00:02:15.240 --> 00:02:18.150 we can rely on the vendors 45 00:02:18.150 --> 00:02:21.630 and the model's function calling capabilities, 46 00:02:21.630 --> 00:02:25.650 and the model is going to return which function to call, 47 00:02:25.650 --> 00:02:28.710 and it's going to do a very good job doing it. 48 00:02:28.710 --> 00:02:31.470 And the function to call, 49 00:02:31.470 --> 00:02:32.820 it's going to specify it 50 00:02:32.820 --> 00:02:35.850 in a very special place in the request. 51 00:02:35.850 --> 00:02:40.680 So we do not need this special ReAct prompt anymore 52 00:02:40.680 --> 00:02:43.920 and this weird output parsing that we did, 53 00:02:43.920 --> 00:02:46.770 which was really hard but not that reliable. 54 00:02:46.770 --> 00:02:49.170 So now we shift everything to the vendor, 55 00:02:49.170 --> 00:02:51.540 the vendor now is responsible to do that, 56 00:02:51.540 --> 00:02:53.310 and we get the information 57 00:02:53.310 --> 00:02:57.960 about which function to call in the LLM's response. 58 00:02:57.960 --> 00:03:02.400 Now, we solved an earlier problem of this output parsing, 59 00:03:02.400 --> 00:03:03.660 which was not reliable, 60 00:03:03.660 --> 00:03:05.730 but we introduced now a new problem 61 00:03:05.730 --> 00:03:08.670 because every vendor did what they want, 62 00:03:08.670 --> 00:03:11.700 and the information about which function to call 63 00:03:11.700 --> 00:03:15.270 was placed differently in different places. 64 00:03:15.270 --> 00:03:17.550 So one called it function calling, 65 00:03:17.550 --> 00:03:19.050 the other tool calling, 66 00:03:19.050 --> 00:03:22.050 it was in different parts of the response. 67 00:03:22.050 --> 00:03:26.010 So what LangChain did is create one single interface 68 00:03:26.010 --> 00:03:29.340 that is going to be called the tool calling interface, 69 00:03:29.340 --> 00:03:33.210 and it implemented all of the integrations to the vendor. 70 00:03:33.210 --> 00:03:36.540 So now we have only one interface of tool calling, 71 00:03:36.540 --> 00:03:40.410 and it's going to get all of the functions to call 72 00:03:40.410 --> 00:03:41.910 and the information about it 73 00:03:41.910 --> 00:03:44.130 and it's going to work for every vendor. 74 00:03:44.130 --> 00:03:46.200 So this was very useful. 75 00:03:46.200 --> 00:03:49.200 Now, when LangGraph came out, 76 00:03:49.200 --> 00:03:53.670 it had a completely different architectural approach. 77 00:03:53.670 --> 00:03:58.200 So instead of having a function-based agent loop, 78 00:03:58.200 --> 00:04:01.710 which was really a while loop abstracted 79 00:04:01.710 --> 00:04:04.890 with the agent executor class of LangChain. 80 00:04:04.890 --> 00:04:07.710 So there wasn't much flexibility 81 00:04:07.710 --> 00:04:09.930 and there wasn't much visibility 82 00:04:09.930 --> 00:04:13.830 or any way to control this loop here. 83 00:04:13.830 --> 00:04:16.710 So then LangGraph came out, 84 00:04:16.710 --> 00:04:21.690 and LangGraph modeled agents as graphs. 85 00:04:21.690 --> 00:04:24.750 So they had nodes, they had edges, 86 00:04:24.750 --> 00:04:28.260 and they had between the nodes a shared state. 87 00:04:28.260 --> 00:04:32.040 So the LangGraph ReAct agent had three components. 88 00:04:32.040 --> 00:04:34.560 It had the state, which was a dictionary 89 00:04:34.560 --> 00:04:36.780 that maintained the conversation, 90 00:04:36.780 --> 00:04:40.050 maybe some intermediate results, 91 00:04:40.050 --> 00:04:43.920 and nodes are simple Python functions 92 00:04:43.920 --> 00:04:45.870 that are receiving the state, 93 00:04:45.870 --> 00:04:47.610 perform the computation, 94 00:04:47.610 --> 00:04:50.550 like maybe to call the LLM or to execute the tool, 95 00:04:50.550 --> 00:04:53.490 and then they return an updated state. 96 00:04:53.490 --> 00:04:56.610 And the edges, they define the control flow. 97 00:04:56.610 --> 00:04:59.820 And by the way, the great motivation for LangGraph 98 00:04:59.820 --> 00:05:04.560 is that LangChain really saw that in every paper 99 00:05:04.560 --> 00:05:08.280 in almost every time you describe an agent, 100 00:05:08.280 --> 00:05:13.110 you actually describe a graph with nodes and with edges. 101 00:05:13.110 --> 00:05:16.530 So this architecture change really helped 102 00:05:16.530 --> 00:05:19.080 to unhide the control flow 103 00:05:19.080 --> 00:05:22.590 because now we have an explicit graph structure 104 00:05:22.590 --> 00:05:25.650 that we can even print and show us a picture, 105 00:05:25.650 --> 00:05:28.350 and it's really hard to understand what's happening 106 00:05:28.350 --> 00:05:29.940 and what's being executed. 107 00:05:29.940 --> 00:05:33.480 Now, if we wanted in the old agent executor 108 00:05:33.480 --> 00:05:35.280 to have some kind of state, 109 00:05:35.280 --> 00:05:36.900 it was really hard to do 110 00:05:36.900 --> 00:05:39.450 and we had to do it with keyword arguments 111 00:05:39.450 --> 00:05:41.010 with a config object, 112 00:05:41.010 --> 00:05:43.230 and it was actually very hard to do. 113 00:05:43.230 --> 00:05:46.320 But with LangGraph and the state schema 114 00:05:46.320 --> 00:05:47.700 is actually very easy to do. 115 00:05:47.700 --> 00:05:50.160 We simply defined in the state 116 00:05:50.160 --> 00:05:53.100 what is the field we want to keep track of. 117 00:05:53.100 --> 00:05:55.200 So if we wanted to know what's happening 118 00:05:55.200 --> 00:05:57.270 with our agent execution 119 00:05:57.270 --> 00:06:00.270 in the agent executor original version, 120 00:06:00.270 --> 00:06:02.100 we really didn't have a way to do it. 121 00:06:02.100 --> 00:06:05.190 So we didn't have a way to monitor and to trace 122 00:06:05.190 --> 00:06:07.110 and to keep track of what's happening. 123 00:06:07.110 --> 00:06:10.080 But with LangGraph, we have automatic checkpoints. 124 00:06:10.080 --> 00:06:12.600 So every time before we execute a node, 125 00:06:12.600 --> 00:06:14.790 LangChain is going to persist the state 126 00:06:14.790 --> 00:06:16.860 and we can have access to that 127 00:06:16.860 --> 00:06:19.860 so we can see exactly what happened and when. 128 00:06:19.860 --> 00:06:21.480 And we can even rewind 129 00:06:21.480 --> 00:06:23.340 and we can travel back in time 130 00:06:23.340 --> 00:06:25.980 and it gives us much greater flexibility. 131 00:06:25.980 --> 00:06:30.180 Now, the biggest thing that LangGraph unlocked 132 00:06:30.180 --> 00:06:35.010 is the fact that we can now compose graphs, 133 00:06:35.010 --> 00:06:38.400 which are agents, one inside another. 134 00:06:38.400 --> 00:06:42.720 So we can actually use a LangGraph graph 135 00:06:42.720 --> 00:06:44.790 as a LangGraph node, 136 00:06:44.790 --> 00:06:46.920 and this is very convenient for us. 137 00:06:46.920 --> 00:06:49.560 And it comes out-of-the-box tracing 138 00:06:49.560 --> 00:06:52.650 and all of the benefits that LangGraph gives us. 139 00:06:52.650 --> 00:06:55.200 And this is something which was very hard to do 140 00:06:55.200 --> 00:06:57.720 with the original agent executor. 141 00:06:57.720 --> 00:07:00.150 So this LangGraph agent was built 142 00:07:00.150 --> 00:07:02.250 under the LangGraph library, 143 00:07:02.250 --> 00:07:04.950 under the graph prebuilt agents 144 00:07:04.950 --> 00:07:07.380 and lived there for a while. 145 00:07:07.380 --> 00:07:10.980 We actually implemented a very similar version ourselves 146 00:07:10.980 --> 00:07:12.270 in this section. 147 00:07:12.270 --> 00:07:17.270 And then LangChain and LangGraph reached version 1.0 148 00:07:17.730 --> 00:07:21.030 and it brought a cleaner API service 149 00:07:21.030 --> 00:07:23.970 with the new create_agent function, 150 00:07:23.970 --> 00:07:27.630 and LangChain then deprecated 151 00:07:27.630 --> 00:07:31.290 and replaced the Create ReAct agent 152 00:07:31.290 --> 00:07:33.540 and the LangGraph prebuilt agent 153 00:07:33.540 --> 00:07:37.260 and simply put everything under this create_agent function, 154 00:07:37.260 --> 00:07:40.560 which returns a compiled graph, 155 00:07:40.560 --> 00:07:42.750 which is LangGraph under the hood, 156 00:07:42.750 --> 00:07:46.167 but with a simple interface for people to use it. 157 00:07:46.167 --> 00:07:47.520 And in this interface, 158 00:07:47.520 --> 00:07:50.610 we can simply give it to the models and the tools, 159 00:07:50.610 --> 00:07:54.450 and boom, we get a ReAct agent which is ready to go, 160 00:07:54.450 --> 00:07:58.620 which has all of the goodies that LangGraph can offer us 161 00:07:58.620 --> 00:08:01.410 with observability, with debugging. 162 00:08:01.410 --> 00:08:03.900 And it also has more flexibility 163 00:08:03.900 --> 00:08:08.400 because we can actually go and customize it ourselves. 164 00:08:08.400 --> 00:08:11.700 All righty, so I hope you enjoyed this video, 165 00:08:11.700 --> 00:08:14.970 and I hope you enjoyed the flow of learning 166 00:08:14.970 --> 00:08:17.970 in general of LangChain agents. 167 00:08:17.970 --> 00:08:20.730 And I think it's really, really important 168 00:08:20.730 --> 00:08:23.280 to know how everything started. 169 00:08:23.280 --> 00:08:26.580 And that you now know, of course, 170 00:08:26.580 --> 00:08:29.910 how to use LangChain create_agent, 171 00:08:29.910 --> 00:08:33.690 but you know exactly how it's implemented under the hood. 172 00:08:33.690 --> 00:08:35.580 And for you, it's not magic, 173 00:08:35.580 --> 00:08:37.470 you know exactly how it's implemented. 174 00:08:37.470 --> 00:08:40.440 And in fact, this interface is just an easier way 175 00:08:40.440 --> 00:08:41.760 for you to use it. 176 00:08:41.760 --> 00:08:46.050 And it's really the foundation for modern LLM agents 177 00:08:46.050 --> 00:08:47.130 and deep agents, 178 00:08:47.130 --> 00:08:49.173 which I hope to cover in the course.