WEBVTT 00:00.400 --> 00:04.320 Okay, now it's on to a proper example that we'll use with Landgraaf. 00:04.680 --> 00:05.440 Just take the obvious. 00:05.440 --> 00:10.040 The reason I showed you the silly example was because I wanted to show that nodes don't need to have 00:10.040 --> 00:13.320 calls to llms, and they still do what they're meant to do. 00:13.360 --> 00:15.360 But now we are going to add an LLM. 00:15.400 --> 00:17.880 So we start by defining the state. 00:18.080 --> 00:21.040 We create a graph builder with that state. 00:21.160 --> 00:26.480 And now we create an LLM a real LLM using chat OpenAI. 00:26.520 --> 00:31.840 So chat OpenAI is a construct from Lang Chain the sibling to Landgraf. 00:32.000 --> 00:35.520 Uh and that's what we'll be using to connect with our LLM. 00:35.680 --> 00:39.640 And now you don't need to use lang chains llms for this. 00:39.640 --> 00:41.120 You can use any Llms. 00:41.120 --> 00:43.160 You could directly call the LLM yourself. 00:43.320 --> 00:47.560 You could also, uh, use maybe OpenAI agents SDK. 00:47.840 --> 00:53.320 But it does make things a bit simpler sometimes if you use Lang chain and most of the community examples 00:53.320 --> 00:55.840 of course go from Landgraf to Lang chain. 00:55.840 --> 00:57.520 So it's it's easy to do it that way. 00:57.520 --> 00:58.880 And that's what we'll do for here. 00:59.440 --> 01:03.150 So we're going to create a new node called chatbot node. 01:03.150 --> 01:06.070 It takes an old state and it returns a new state. 01:06.230 --> 01:07.390 And what does it do. 01:07.430 --> 01:12.390 Well it takes the the LM and it invokes on that LM. 01:12.390 --> 01:13.830 So it's again it's the Lang chain. 01:13.830 --> 01:18.950 Landgraf word invoke uh, passing in the messages from old state. 01:18.950 --> 01:21.950 So old state has a messages field and that is what we pass in. 01:22.270 --> 01:29.710 And then for the new state, it creates a new state object which contains within it as in its messages 01:29.710 --> 01:31.790 field, it contains the response. 01:31.790 --> 01:34.230 And we return the new state. 01:34.510 --> 01:40.310 And we add that node called chatbot, uh, into our graph builder. 01:40.350 --> 01:40.830 Done. 01:41.190 --> 01:47.150 Now we'll add some edges from start to chatbot from chatbot to end done. 01:47.510 --> 01:49.710 And now we will compile our graph. 01:49.710 --> 01:51.510 Step five and we'll look at the graph. 01:51.510 --> 01:54.750 And sure enough it goes start to chatbot to end. 01:55.350 --> 01:59.070 And then we put it all together in a simple gradio chat function. 01:59.070 --> 02:05.260 It takes an initial state, which is a state object set up with these messages like so. 02:05.700 --> 02:09.620 We then call graph dot invoke to actually call our graph. 02:09.740 --> 02:16.260 We print the result and we will also show the results back in Gradio. 02:16.540 --> 02:17.460 So here it is. 02:17.500 --> 02:18.860 And I can say hi there. 02:20.300 --> 02:22.180 And it's actually calling OpenAI. 02:22.220 --> 02:25.260 Now it's not using our silly adjectives. 02:25.420 --> 02:32.540 And you'll see down here that there's the user message and the response coming in these objects, this 02:32.540 --> 02:35.460 human message object that is coming back. 02:35.460 --> 02:36.780 Sorry, that human message is my. 02:36.780 --> 02:37.340 Hi there. 02:37.380 --> 02:38.980 I mean, there should be an AI message. 02:38.980 --> 02:39.740 There it is. 02:39.940 --> 02:44.740 There is the AI message, which is the response actually coming from OpenAI. 02:45.420 --> 02:50.860 But one thing that's worth noting is that if I continue this conversation every time we are invoking 02:50.980 --> 02:56.660 this graph, and you will see what you have probably already suspected, which is that we're not actually 02:56.700 --> 02:58.740 keeping track of any history here. 02:58.970 --> 03:00.410 Let's see that in action. 03:00.610 --> 03:02.770 Uh, if I, uh, say, um. 03:03.010 --> 03:03.890 My name's Ed. 03:07.490 --> 03:08.410 Nice to meet you, Ed. 03:08.450 --> 03:09.770 How can I assist you today? 03:11.290 --> 03:12.290 What's my name? 03:15.050 --> 03:17.770 I'm sorry, but I don't have access to your personal data. 03:17.970 --> 03:21.450 So there's a sign that it's not able to keep context. 03:21.450 --> 03:26.370 And you can see it yourself if you read the information that's going to and fro. 03:26.530 --> 03:33.050 So because we've just got this simple graph that we were invoking each time, there's nothing particularly 03:33.050 --> 03:34.770 interesting happening here. 03:34.810 --> 03:41.370 And the state uh, is just, just contains that that uh, doesn't contain the history or anything. 03:41.610 --> 03:44.850 So, uh, that's one of the things that we clearly need to address. 03:45.170 --> 03:47.370 And the good news is that we will indeed address it. 03:47.370 --> 03:49.450 But the bad news is not until tomorrow. 03:49.570 --> 03:56.170 But we'll also address things like tools, our old favorite, along with a couple of other things. 03:56.330 --> 03:58.010 So, uh, look forward to it. 03:58.010 --> 03:58.810 I'll see you then.