WEBVTT 00:00.000 --> 00:01.200 -: In this video we're gonna have a look at 00:01.200 --> 00:03.300 how you can customize LangGraph state 00:03.300 --> 00:05.490 beyond just using the messages key 00:05.490 --> 00:07.020 inside of the state graph. 00:07.020 --> 00:08.220 This is important because 00:08.220 --> 00:11.580 what you wanna do is not just have specifically 00:11.580 --> 00:14.220 the message history, but if you can have lots of other keys 00:14.220 --> 00:16.890 inside LangGraph, that's gonna give you the flexibility 00:16.890 --> 00:19.050 to have lots of different types of Boolean flags 00:19.050 --> 00:22.590 so that you can control different types of edges and nodes 00:22.590 --> 00:25.020 and conditional edges based on some Booleans 00:25.020 --> 00:26.550 or some state that's stored 00:26.550 --> 00:28.710 inside of the LangGraph state machine. 00:28.710 --> 00:30.480 We've got some imports that we do 00:30.480 --> 00:33.300 and we're also just setting up my Tavily key, 00:33.300 --> 00:36.030 and then we're gonna have a look at customizing the state. 00:36.030 --> 00:39.420 So again, the important point of this lesson is 00:39.420 --> 00:42.060 that we're doing exactly the same stuff we've done before, 00:42.060 --> 00:45.300 but we've got a new flag, which is to ask_human 00:45.300 --> 00:47.850 and the reason why we are using this ask flag, 00:47.850 --> 00:49.890 and you could have lots of different pieces of state here, 00:49.890 --> 00:53.083 so I could also have ask_for_final_review. 00:54.489 --> 00:56.580 And I could have a Boolean piece of state here, 00:56.580 --> 00:58.650 but we can add lots and lots of different pieces of state. 00:58.650 --> 01:00.750 But the important point is that you can, 01:00.750 --> 01:02.910 if you want to do more complex things outside 01:02.910 --> 01:05.610 of just the message history, then it can be useful 01:05.610 --> 01:08.280 to have extra bits of state inside of LangGraph 01:08.280 --> 01:09.600 that you can manipulate. 01:09.600 --> 01:12.481 So you can see here we have a RequestAssistance 01:12.481 --> 01:16.379 piece of schema and we are going to create a tool, 01:16.379 --> 01:18.829 and the tool that we have is TavilySearchResults. 01:19.778 --> 01:23.280 And we're also going to have this RequestAssistance 01:23.280 --> 01:25.950 as a separate tool that we use. 01:25.950 --> 01:28.187 So you can see where here we have a Pydantic model, 01:28.187 --> 01:30.720 which is escalate the conversation to an expert. 01:30.720 --> 01:33.510 Use this if you're unable to assist directly. 01:33.510 --> 01:35.670 And so we're binding the LLM with both 01:35.670 --> 01:38.093 the Tavily search tool and also the RequestAssistance tool. 01:39.959 --> 01:43.230 And we have a chatbot which will invoke with the tools 01:43.230 --> 01:45.690 and it sets the ask_human to false. 01:45.690 --> 01:47.070 If there are tool calls, 01:47.070 --> 01:51.525 and the tool calls name is the RequestAssistance name, 01:51.525 --> 01:55.339 we set the ask_human to true and then we return the messages 01:55.339 --> 01:57.750 and the ask_human. 01:57.750 --> 02:00.150 Now we then build the state graph 02:00.150 --> 02:02.970 and we have our chatbot node that we had before, 02:02.970 --> 02:04.860 which is now this Python function. 02:04.860 --> 02:06.060 And we have the tools 02:06.060 --> 02:10.080 and the tool node being just the Tavily search results. 02:10.080 --> 02:12.690 Now, when we have a response, 02:12.690 --> 02:14.850 we are also going to have a little helper message 02:14.850 --> 02:16.901 for returning a tool message, 02:16.901 --> 02:19.290 which will just return the response from the AI 02:19.290 --> 02:23.160 and the tool call from the first one inside of that AI. 02:23.160 --> 02:26.460 And we have a human node which is 02:26.460 --> 02:28.020 basically getting a list of messages. 02:28.020 --> 02:29.520 And if the messages, 02:29.520 --> 02:31.230 if there isn't any messages at the moment, 02:31.230 --> 02:33.420 then we don't have a response from the human. 02:33.420 --> 02:36.210 So typically the user will have updated the state 02:36.210 --> 02:37.050 during the interrupt. 02:37.050 --> 02:39.750 If they've chosen not to, we'll include a placeholder, 02:39.750 --> 02:42.750 a tool message just to let the LLM continue. 02:42.750 --> 02:45.360 So that's a nice way of us being able to 02:45.360 --> 02:47.850 always add a tool message on just in case 02:47.850 --> 02:50.210 of the human node doesn't have 02:50.210 --> 02:53.991 a response from the human from the new messages. 02:53.991 --> 02:56.670 But if it does, then what we're gonna do 02:56.670 --> 02:59.220 is then unset the flag to false. 02:59.220 --> 03:01.110 And we have this node now called human. 03:01.110 --> 03:02.160 So we've got a couple of different nodes. 03:02.160 --> 03:04.860 We've got the chatbot node, the tools node, 03:04.860 --> 03:06.750 and also the human node. 03:06.750 --> 03:09.731 And we need a function to define 03:09.731 --> 03:12.060 what is the next node going to be. 03:12.060 --> 03:15.630 So if the ask_humans flag in the LangGraph state 03:15.630 --> 03:19.290 is set to true, we are going to return the human. 03:19.290 --> 03:21.360 And otherwise we're going to have the tool condition, 03:21.360 --> 03:25.173 which will either go to tools or __end__. 03:27.360 --> 03:30.300 We then build our graph with the chatbots 03:30.300 --> 03:31.590 and we select the next node. 03:31.590 --> 03:33.090 And if it's returned to human, 03:33.090 --> 03:35.970 we're gonna go into the human node, 03:35.970 --> 03:37.350 and remembering that the human node 03:37.350 --> 03:39.270 will run this Python function. 03:39.270 --> 03:42.600 And if it's the tools, that will run the tools node, 03:42.600 --> 03:44.580 which we've defined up here. 03:44.580 --> 03:47.070 And remember that the tools node is only using 03:47.070 --> 03:48.840 the Tavily search results. 03:48.840 --> 03:51.690 We have bound two tools to the chat model, 03:51.690 --> 03:53.640 which is the RequestAssistance tool. 03:53.640 --> 03:56.550 But that is only being used not inside 03:56.550 --> 03:58.230 of the tool node directly. 03:58.230 --> 04:01.170 And we also have the end which gets mapped to the end. 04:01.170 --> 04:03.360 And so we've built our conditional edges 04:03.360 --> 04:06.000 and we then add the edges from the tool to the chatbot 04:06.000 --> 04:07.347 and the human to the chatbot. 04:07.347 --> 04:09.330 And we set the entry point to chatbot. 04:09.330 --> 04:12.030 And then we're using a chat pointer here 04:12.030 --> 04:13.740 whilst we compile our graph. 04:13.740 --> 04:16.980 And we are also interrupting before the human node. 04:16.980 --> 04:18.210 And this is important 04:18.210 --> 04:22.080 because we want to update the state of the human node, 04:22.080 --> 04:24.150 but if there wasn't a updated state, 04:24.150 --> 04:25.920 then we would have a placeholder. 04:25.920 --> 04:28.770 But you can assume that before this human node runs, 04:28.770 --> 04:30.300 we are actually gonna update it 04:30.300 --> 04:32.520 so that this one wouldn't happen. 04:32.520 --> 04:34.699 But we would actually have another message 04:34.699 --> 04:36.390 directly already updated. 04:36.390 --> 04:38.880 And then we would go from the human node onwards. 04:38.880 --> 04:41.040 So let's have a look at how this looks. 04:41.040 --> 04:42.360 So you can see this is what we have. 04:42.360 --> 04:44.250 So we've got the chatbot, 04:44.250 --> 04:46.050 which goes into the human if it needs it, 04:46.050 --> 04:47.400 and then goes back into the chatbot. 04:47.400 --> 04:49.290 We also have the tools. 04:49.290 --> 04:50.310 If there are tool calls, 04:50.310 --> 04:52.920 then those will get filtered into the tool node. 04:52.920 --> 04:54.930 And then after that, then it will end. 04:54.930 --> 04:58.320 Now you can say, I need some help guiding for assistance 04:58.320 --> 04:59.790 with an AI agent. 04:59.790 --> 05:01.590 We'll give a thread ID of one 05:01.590 --> 05:03.990 and we'll stream that with the user input. 05:03.990 --> 05:05.010 And let's see what happens. 05:05.010 --> 05:06.840 It's decided it wants to call 05:06.840 --> 05:09.531 the human RequestAssistance tool. 05:09.531 --> 05:14.340 And what we can then do is get the state of the graph 05:14.340 --> 05:17.190 and you can see that the next node is human. 05:17.190 --> 05:19.350 So we've essentially, if we scroll up a bit, 05:19.350 --> 05:21.060 we first enter the chatbot node 05:21.060 --> 05:22.860 and it decided to go to the human node, 05:22.860 --> 05:24.240 but we have interrupted it 05:24.240 --> 05:26.580 before it's gone into the human node. 05:26.580 --> 05:28.860 And then what we are then basically gonna do 05:28.860 --> 05:30.810 is create a tool message. 05:30.810 --> 05:33.150 So we'll get the values, the messages, 05:33.150 --> 05:35.250 the most recent message from the AI. 05:35.250 --> 05:37.020 And you can see this gives us things like 05:37.020 --> 05:40.410 the tool calls that we can get. 05:40.410 --> 05:42.630 And so what we want to do is 05:42.630 --> 05:46.350 we want to use this create_response function 05:46.350 --> 05:48.540 and we'll just have a look at that what that does again. 05:48.540 --> 05:50.310 So that will take the AI message 05:50.310 --> 05:52.770 and it will get the first tool call ID 05:52.770 --> 05:55.410 and it will have a response to that tool. 05:55.410 --> 05:58.020 So if we're gonna scroll down again, 05:58.020 --> 06:00.810 now that we know that it's gonna get the AI message ID 06:00.810 --> 06:03.900 and create a tool message for us, you can see that there. 06:03.900 --> 06:06.240 And then we have the human response, 06:06.240 --> 06:07.442 we are the experts here, 06:07.442 --> 06:09.960 recommend check out our LangGraph to build your agent, 06:09.960 --> 06:13.770 it's much more reliable and flexible and extensible. 06:13.770 --> 06:15.780 And then we're gonna update the state 06:15.780 --> 06:18.090 with the config and the messages. 06:18.090 --> 06:20.670 And we're gonna add on the tool message 06:20.670 --> 06:23.280 that comes back from the create_response. 06:23.280 --> 06:25.440 And now that we've done that, if we get the config, 06:25.440 --> 06:28.290 you'll see that we've also now got a tool message 06:28.290 --> 06:31.500 that has been added on after the AI has decided 06:31.500 --> 06:33.510 that it needed to request assistance. 06:33.510 --> 06:34.890 And so we've almost answered 06:34.890 --> 06:36.570 the request assistance from this. 06:36.570 --> 06:39.870 And therefore if we do a none configuration, 06:39.870 --> 06:41.580 it will resume the graph. 06:41.580 --> 06:44.790 So if we were to do, for example, graph dot, 06:44.790 --> 06:47.110 and then I could do get the state, 06:47.110 --> 06:50.820 and I just wanna prove to you that the next node here 06:50.820 --> 06:53.040 will still be the human node. 06:53.040 --> 06:57.853 But we already know that we've got a input for, 06:57.853 --> 07:00.540 if we just scroll up a bit, we already have an input 07:00.540 --> 07:02.760 for the messages being a tool message 07:02.760 --> 07:05.580 because you've added on a tool message specifically. 07:05.580 --> 07:07.440 And therefore we would skip this step 07:07.440 --> 07:08.820 and we would go straight here 07:08.820 --> 07:09.930 and we're just basically setting 07:09.930 --> 07:11.580 the ask_human flag to be false. 07:11.580 --> 07:14.220 And we've got the messages that are being returned. 07:14.220 --> 07:15.990 So that's basically how it's working 07:15.990 --> 07:18.090 is by adding on this extra tool message. 07:18.090 --> 07:20.400 When we go back into that human node, 07:20.400 --> 07:21.823 which is the next node, 07:21.823 --> 07:23.963 we're actually just gonna return the messages 07:23.963 --> 07:26.670 and then you'll see what happens from there. 07:26.670 --> 07:28.890 So we've got the tool message 07:28.890 --> 07:30.330 and then the AI has said, 07:30.330 --> 07:32.610 experts have recommended using LangGraph 07:32.610 --> 07:34.500 as a reliable and accessible tool. 07:34.500 --> 07:36.780 So then the graph has specifically ended. 07:36.780 --> 07:38.684 So what happened was we went into the chatbot, 07:38.684 --> 07:41.580 we interrupted before the human node, 07:41.580 --> 07:45.300 and then after that what we did was we added on a message, 07:45.300 --> 07:47.070 we then went into the human node, 07:47.070 --> 07:49.080 which went back into the chatbot. 07:49.080 --> 07:51.810 There was now no tools to be called. 07:51.810 --> 07:54.030 And so what we then did is we decided 07:54.030 --> 07:57.090 to end the LangGraph state machine. 07:57.090 --> 08:00.403 So, cool, so that's a bit of what we've done there. 08:00.403 --> 08:01.890 So hopefully that made sense. 08:01.890 --> 08:03.810 The important point here is you want 08:03.810 --> 08:05.520 to have your interrupt befores 08:05.520 --> 08:09.390 or interrupt afters as you build your state machine. 08:09.390 --> 08:12.450 And you could also have done it with an interrupt after. 08:12.450 --> 08:14.040 The important point here is if you did do 08:14.040 --> 08:16.200 an interrupt after, you would need to make sure 08:16.200 --> 08:18.120 that the node kind of handles for the fact 08:18.120 --> 08:21.960 that you wouldn't have that after inside the node. 08:21.960 --> 08:23.340 There's various different strategies. 08:23.340 --> 08:24.780 One is interrupting before the node 08:24.780 --> 08:26.760 and then resuming from there. 08:26.760 --> 08:28.170 One is interrupting after 08:28.170 --> 08:30.163 and then updating the previous bit of state. 08:30.163 --> 08:33.540 And again, we're using a checkpoint to which allows us 08:33.540 --> 08:34.950 to run different threads. 08:34.950 --> 08:38.160 And LangGraph is clever enough to allow us 08:38.160 --> 08:40.705 to specifically customize the state 08:40.705 --> 08:43.680 and we can use that customizable state 08:43.680 --> 08:46.020 to run different types of conditional edges. 08:46.020 --> 08:48.720 Cool, so in the next one we're gonna have a look at 08:48.720 --> 08:50.550 a concept called time travel. 08:50.550 --> 08:54.150 Time travel will allow you to specifically wind 08:54.150 --> 08:57.930 forwards or backwards graph state inside of LangGraph 08:57.930 --> 09:00.150 from using a check pointer. 09:00.150 --> 09:02.280 This can be useful if you want users to be able 09:02.280 --> 09:04.865 to undo changes they've had or revert changes, 09:04.865 --> 09:07.620 and having the version control over the graph 09:07.620 --> 09:09.630 as well as being able to update the graph. 09:09.630 --> 09:11.380 Cool, I'll see you in the next one.