WEBVTT 00:00.150 --> 00:01.380 -: Hey, welcome back. 00:01.380 --> 00:02.910 And in this video we're gonna have a look at 00:02.910 --> 00:05.250 how you can use tools, 00:05.250 --> 00:08.190 and also persistence, inside of LangGraph. 00:08.190 --> 00:10.380 Firstly, we're gonna just import a couple of packages 00:10.380 --> 00:13.200 and then after that we are also going 00:13.200 --> 00:15.930 to see if you have your open AI, API key 00:15.930 --> 00:17.310 specifically for that, as well. 00:17.310 --> 00:19.470 So I'm just gonna run this Python. 00:19.470 --> 00:21.150 And the other thing we're gonna need 00:21.150 --> 00:22.608 is something called Tavily. 00:22.608 --> 00:25.440 And Tavily is a great search engine that you can use, 00:25.440 --> 00:28.560 they offer a free amount specifically of credits. 00:28.560 --> 00:31.140 And the main benefit of using Tavily is 00:31.140 --> 00:34.170 that you can easily connect your large language models 00:34.170 --> 00:37.290 with text-based results that come back from Google search. 00:37.290 --> 00:40.200 So, rather than you having to use a Google search, 00:40.200 --> 00:43.290 API wrapper and get the text from the webpages, 00:43.290 --> 00:45.030 all of that is handled automatically 00:45.030 --> 00:47.970 for you using a package like Tavily. 00:47.970 --> 00:50.010 So that's also one of the requirements. 00:50.010 --> 00:53.190 And then I'm also gonna set my API key for Tavily. 00:53.190 --> 00:55.201 So I'm just gonna copy that in 00:55.201 --> 00:58.290 and then we're gonna use the Tavily results, 00:58.290 --> 01:01.440 and you can see this is the tool that we'll be using, 01:01.440 --> 01:04.410 and we've set a max results of two 01:04.410 --> 01:06.150 and then we wrap that in a Python list 01:06.150 --> 01:07.440 'cause we're gonna use a list of tools 01:07.440 --> 01:09.437 and you can even invoke a tool. 01:09.437 --> 01:12.060 So you can say, like, "What's a node in LangGraph?" 01:12.060 --> 01:13.830 And you'll see that it gives us a couple 01:13.830 --> 01:15.357 of different results and you've got the URL 01:15.357 --> 01:18.153 and the content specifically of those. 01:19.020 --> 01:21.030 Now, the next thing that we need to do 01:21.030 --> 01:23.460 is we need to annotate our state. 01:23.460 --> 01:26.580 And you can see we have a state graph being set up 01:26.580 --> 01:27.750 and that is being set up 01:27.750 --> 01:30.750 using this state-typed dictionary we have, 01:30.750 --> 01:32.340 which has a list of messages 01:32.340 --> 01:34.380 which are also being appended to using 01:34.380 --> 01:37.260 that add messages that will merge in the left 01:37.260 --> 01:39.390 and right messages easily. 01:39.390 --> 01:40.860 So that's a nice helper function 01:40.860 --> 01:43.590 and we're gonna be using the GPT-4 O model 01:43.590 --> 01:46.140 and then also making sure that we've got these tools. 01:46.140 --> 01:49.620 So we have a list of tools that you can see 01:49.620 --> 01:51.300 with the Tavily search results, 01:51.300 --> 01:54.000 and What we're gonna do is then take our chat model, 01:54.000 --> 01:56.730 which is this LLM, and we're gonna bind those tools to it. 01:56.730 --> 01:59.520 And different types of providers like Anthropic 01:59.520 --> 02:03.150 and Grok will also provide you with the ability 02:03.150 --> 02:06.330 to add tools to your large language models. 02:06.330 --> 02:07.920 We've got a function here called chatbot, 02:07.920 --> 02:10.530 which is basically what we had before, 02:10.530 --> 02:13.140 and it returns a key of messages 02:13.140 --> 02:16.770 with the invoked state of the messages with the tools. 02:16.770 --> 02:18.470 So we're gonna add that as a node. 02:19.470 --> 02:20.640 Now we need to create a function 02:20.640 --> 02:22.563 to run the tools if they are called, 02:23.910 --> 02:25.590 just to show you that the sort of 02:25.590 --> 02:27.960 the lower level implementation, 02:27.960 --> 02:30.810 remember that your chat bot can return 02:30.810 --> 02:34.260 that it wants to call several tools in one message. 02:34.260 --> 02:36.090 And so what we need is something 02:36.090 --> 02:37.770 that will have all of the tools 02:37.770 --> 02:39.750 with the names of those tools 02:39.750 --> 02:43.710 and we also have a call here which basically goes 02:43.710 --> 02:45.000 and tries to get the messages. 02:45.000 --> 02:48.600 And if there aren't any messages then it will raise an error 02:48.600 --> 02:50.460 and it makes the outputs, 02:50.460 --> 02:53.910 and for each message doc tool calls, 02:53.910 --> 02:55.170 it will get the tool call, 02:55.170 --> 02:57.840 and it will find that tool by its name 02:57.840 --> 03:02.400 and invoke it with a tool call with the ARGs of that 03:02.400 --> 03:05.910 and it will append the outputs as tool messages 03:05.910 --> 03:09.720 with the content from the json.dumps is the content 03:09.720 --> 03:12.060 and the name and the tool call ID. 03:12.060 --> 03:14.550 The tool call IDs are quite important 03:14.550 --> 03:16.170 because that tells that this, you know, 03:16.170 --> 03:19.080 that what message this tool call is responding to. 03:19.080 --> 03:21.690 And then we return the messages with the outputs 03:21.690 --> 03:24.300 and then you can see we've got a basic tool node 03:24.300 --> 03:26.190 which takes in a bunch of tools 03:26.190 --> 03:28.920 and we can use our original graph builder 03:28.920 --> 03:31.020 to add on this new tool node. 03:31.020 --> 03:33.570 Now once, because we've got two nodes, 03:33.570 --> 03:35.940 now we've got the chat bot node and the tools node, 03:35.940 --> 03:39.330 the next thing we have to do is we have to link these things 03:39.330 --> 03:42.750 and recall that edges are the way that we can connect nodes 03:42.750 --> 03:43.920 inside of a state machine. 03:43.920 --> 03:45.960 So, if we're thinking about a state machine 03:45.960 --> 03:48.480 and how these kinds of nodes, you know, might look, 03:48.480 --> 03:50.520 you might have something that looks a little bit like this 03:50.520 --> 03:52.800 where you've got H for Human 03:52.800 --> 03:55.050 and LLM for Large Language Model 03:55.050 --> 03:57.480 and you could have some types of nodes 03:57.480 --> 04:00.180 where you've got these sort of bubbles and black boxes 04:00.180 --> 04:04.020 and you can see the arrows pointing to the directional flow. 04:04.020 --> 04:06.870 And we can think of these black arrows 04:06.870 --> 04:09.150 in the middle here as the edges. 04:09.150 --> 04:11.730 Okay? And the great thing about LangGraph 04:11.730 --> 04:14.910 is we can have these cyclical loops where, you know, 04:14.910 --> 04:17.070 and we can also have conditional edges. 04:17.070 --> 04:19.110 So, a conditional edge might look like this 04:19.110 --> 04:21.240 where we have a node here 04:21.240 --> 04:25.260 and we decide whether we should progress to this next one. 04:25.260 --> 04:27.720 So I'll do blue for conditional. 04:27.720 --> 04:29.160 We might progress here, 04:29.160 --> 04:33.300 or we might go and we might go and do something else 04:33.300 --> 04:35.490 and then go back into here, right? 04:35.490 --> 04:37.890 So this might be something that we do. 04:37.890 --> 04:39.870 So, you can think of it like a little bit like this, 04:39.870 --> 04:42.920 so, we might go into here, 04:42.920 --> 04:45.120 so this is kind of what I'm thinking. 04:45.120 --> 04:47.040 So we've got a node 04:47.040 --> 04:52.040 and we might conditionally go into this node. 04:52.920 --> 04:54.766 You can think of this as the conditional node. 04:54.766 --> 04:57.085 (keyboard keys rapidly clicking) 04:57.085 --> 04:59.843 And we can think of this as if we've passed 05:01.180 --> 05:03.330 a certain check continue, right? 05:03.330 --> 05:05.250 And we'll go to the continue node. 05:05.250 --> 05:09.540 And what's important about this is we essentially have 05:09.540 --> 05:14.010 between this node here and this continue node, 05:14.010 --> 05:17.430 we have a conditional edge, right? 05:17.430 --> 05:20.670 Sometimes we're going back in, sometimes we're going here. 05:20.670 --> 05:23.580 So, this edge is fairly conditional 05:23.580 --> 05:24.870 and you can think of it like maybe 05:24.870 --> 05:27.150 you will have two different conditions. 05:27.150 --> 05:30.090 One goes to the blue and one goes to the green 05:30.090 --> 05:31.080 and decides to continue. 05:31.080 --> 05:32.610 So that's what we're gonna have a look at now. 05:32.610 --> 05:35.760 It's pretty simple, it's just mapping, you know, 05:35.760 --> 05:37.680 this node to that node based on a condition 05:37.680 --> 05:39.580 and this node to the conditional node, 05:41.697 --> 05:42.530 based on a condition, as well. 05:42.530 --> 05:43.620 So, let's have a look at that. 05:43.620 --> 05:46.560 So we've got this route tools function 05:46.560 --> 05:51.180 and you can see we are returning a string of literal, 05:51.180 --> 05:53.670 so basically just a string of Veeva tools 05:53.670 --> 05:58.080 or this underscore, underscore, end, underscore, underscore. 05:58.080 --> 06:02.160 Now, what you'll see is if we have a list, 06:02.160 --> 06:04.200 get the AI message. 06:04.200 --> 06:08.550 If the messages exist there, we're gonna get that instead. 06:08.550 --> 06:12.060 And if we don't have any of those, we raise an error, 06:12.060 --> 06:13.837 and we basically have this bit where we say, 06:13.837 --> 06:18.000 "If this AI message has tool calls, 06:18.000 --> 06:20.070 and the tool calls is greater than zero, 06:20.070 --> 06:23.700 I want to go into this tools node. 06:23.700 --> 06:27.120 Else, I want to just return with this end." 06:27.120 --> 06:29.040 Now the end is a very special node 06:29.040 --> 06:32.310 because it represents the end of the state machine. 06:32.310 --> 06:35.913 And what we then do is with our route tools function, 06:36.810 --> 06:40.020 we are then going to map the route tools 06:40.020 --> 06:41.250 with the outputs of these. 06:41.250 --> 06:42.840 So you can see we've got the return of tools 06:42.840 --> 06:47.280 and the return of end, the return of tools gets matched 06:47.280 --> 06:50.760 to our tools node that we created earlier, right? 06:50.760 --> 06:52.980 And I'll just scroll up to just remind you, 06:52.980 --> 06:55.440 this is where we created the tools node, right? 06:55.440 --> 06:57.390 So if I was to make this more explicit 06:57.390 --> 06:59.340 I would call this the tools node 06:59.340 --> 07:01.530 and then if I was to go down here, 07:01.530 --> 07:04.950 then this tools would then map to the tools node here. 07:04.950 --> 07:05.783 Okay? 07:07.620 --> 07:09.720 It's this bit which is actually mapping it. 07:09.720 --> 07:10.650 'Cause you can think of it, like, 07:10.650 --> 07:12.540 we have this route tools function, 07:12.540 --> 07:16.020 the outputs of these are the keys inside of the dictionary. 07:16.020 --> 07:17.281 So, tools and end, 07:17.281 --> 07:21.330 which get mapped to either tools or tools node, 07:21.330 --> 07:22.800 depending upon what you decided 07:22.800 --> 07:24.480 to call your tools node. 07:24.480 --> 07:27.390 So, now that I've just gone and update that 07:27.390 --> 07:29.580 and I'll just run this and, yeah, 07:29.580 --> 07:31.980 this has given me an error 'cause that node already exists, 07:31.980 --> 07:33.120 which is fine, 07:33.120 --> 07:35.730 then what we can do is we can compile this graph 07:35.730 --> 07:38.010 and this probably already gives us an error 07:38.010 --> 07:40.050 because we've already compiled the graph. 07:40.050 --> 07:42.300 So ,you can only run these types of things once 07:42.300 --> 07:44.040 because, otherwise, you will get an error. 07:44.040 --> 07:47.430 But if we have a look at our graph, you have the start, 07:47.430 --> 07:50.460 we've set the entry point as the chat bot, 07:50.460 --> 07:54.510 and we've got the different types of of nodes. 07:54.510 --> 07:56.880 So we could also have a look at the nodes, 07:56.880 --> 07:59.430 and you can see we have a start node, 07:59.430 --> 08:02.790 we have a chat bot node and we have a tool node. 08:02.790 --> 08:06.870 Yeah, and the conditional edges is basically going 08:06.870 --> 08:08.040 to do something like this. 08:08.040 --> 08:10.410 If we look, so we'll start here, 08:10.410 --> 08:12.360 we'll go into the chat bot node, 08:12.360 --> 08:14.760 if there are tools, we will go and do those, 08:14.760 --> 08:16.110 and we'll go back into chat bot. 08:16.110 --> 08:18.150 And if there are no more tool calls, 08:18.150 --> 08:20.883 then we will also then just end the chat bot. 08:21.750 --> 08:23.430 And now we can ask the bot questions 08:23.430 --> 08:25.110 outside of its training data. 08:25.110 --> 08:27.450 So, I've got a little while true here 08:27.450 --> 08:30.450 and we have, if the user enters quit or exit or queue, 08:30.450 --> 08:32.490 it will break and print goodbye. 08:32.490 --> 08:35.160 Otherwise, it will stream the graph with the messages 08:35.160 --> 08:37.260 of a two-pull object in the list 08:37.260 --> 08:38.880 where you've got the user and the user input. 08:38.880 --> 08:41.280 And we're just printing out the event values 08:41.280 --> 08:43.080 inside of the LangGraph. 08:43.080 --> 08:44.527 So if I now run this and say, 08:44.527 --> 08:46.650 "Hey, how are you?" (keyboard keys tapping) 08:46.650 --> 08:48.060 There aren't any tool calls here. 08:48.060 --> 08:50.040 So, the assistant has basically gone straight 08:50.040 --> 08:53.190 from the chat bot node directly to the end node, right? 08:53.190 --> 08:56.340 So, there's a very simple architecture we have 08:56.340 --> 08:57.930 here with this conditional edge 08:57.930 --> 08:59.850 where we're either going into the tools node 08:59.850 --> 09:01.170 or we're going to the end node 09:01.170 --> 09:03.547 after we've invoked the model so it can say, 09:03.547 --> 09:04.909 "I am good, thanks. (keyboard keys clicking) 09:04.909 --> 09:08.250 What tools do you have available?" 09:08.250 --> 09:10.920 And this probably also won't call a tool. 09:10.920 --> 09:13.710 It says we've got Tavily search 09:13.710 --> 09:16.560 and we also have multi-tool use. 09:16.560 --> 09:20.631 Now, one thing we might wanna say is, "Can you search 09:20.631 --> 09:24.540 on Google for digital marketing leads?" 09:24.540 --> 09:28.050 And then what it's gonna do is it's then gonna go through 09:28.050 --> 09:31.500 and it's gonna use this chat bot and use the tools 09:31.500 --> 09:33.000 and then it's gonna go back here 09:33.000 --> 09:35.160 and if there aren't any more tools to call, 09:35.160 --> 09:36.660 it's then gonna end. 09:36.660 --> 09:38.610 And so what you can see here is it decided 09:38.610 --> 09:41.490 to call Tavily search here, 09:41.490 --> 09:44.460 and then it also then provided the actual answer here. 09:44.460 --> 09:46.500 So it gave us a marketing deep dive 09:46.500 --> 09:48.030 and so that's basically it. 09:48.030 --> 09:50.730 So, congratulations, you've figured out how 09:50.730 --> 09:53.970 to add a tool in a way which works inside of LangGraph, 09:53.970 --> 09:55.420 which is really, really cool. 09:56.400 --> 09:57.840 Now the next thing that we wanna cover 09:57.840 --> 09:59.700 is adding memory to this chat bot. 09:59.700 --> 10:01.560 So we've got a conditional edge, 10:01.560 --> 10:03.390 which will either call the tool, 10:03.390 --> 10:05.460 or it will just end the chat bot. 10:05.460 --> 10:07.080 Now, the next thing we have to look at 10:07.080 --> 10:10.230 is how do we specifically save the state 10:10.230 --> 10:12.090 of which our graph is currently in? 10:12.090 --> 10:14.610 This is important because you might want to 10:14.610 --> 10:16.170 save the state of the graph, 10:16.170 --> 10:18.150 you might want to reload the graph, 10:18.150 --> 10:20.190 you might wanna update parts of that graph, 10:20.190 --> 10:22.620 and then, you know, getting some human feedback. 10:22.620 --> 10:24.630 And, so, where you can use this concept 10:24.630 --> 10:26.250 called a check pointer, 10:26.250 --> 10:30.390 and we're using the thread ID as the unique identifier 10:30.390 --> 10:32.790 for a specific type of graph state, 10:32.790 --> 10:35.730 which will be automatically saved to a database. 10:35.730 --> 10:38.640 Now, we're using the SQL lite saver 10:38.640 --> 10:41.760 and we're deciding to store that directly inside of memory, 10:41.760 --> 10:43.770 but there are different types of persistence 10:43.770 --> 10:44.820 that you might want to know about. 10:44.820 --> 10:47.550 So there is a persistence notebook 10:47.550 --> 10:49.590 inside of LangChain's documentation 10:49.590 --> 10:51.690 that provides for PostgreSQL. 10:51.690 --> 10:54.000 So I'd recommend having a look at that 10:54.000 --> 10:56.310 and I'll leave a link in this lecture 10:56.310 --> 10:59.940 for how they've implemented the PostgreSQL check pointer 10:59.940 --> 11:02.700 for if you want to use that for your production systems. 11:02.700 --> 11:04.350 Let's have a look at how we could do this 11:04.350 --> 11:06.600 purely using the memory check pointer. 11:06.600 --> 11:10.440 And we are then going to replace the basic tool node 11:10.440 --> 11:14.220 with LangChain's built-in tool node and tools condition, 11:14.220 --> 11:15.990 which basically do the exact same things 11:15.990 --> 11:17.763 with the parallel tool cooling, 11:18.630 --> 11:19.950 but it's basically wrapped up 11:19.950 --> 11:21.780 and you don't have to worry about that. 11:21.780 --> 11:23.910 So, you can see here we're now importing 11:23.910 --> 11:25.470 a couple of prebuilt things. 11:25.470 --> 11:26.940 So you've got this tool node 11:26.940 --> 11:29.090 that basically does the same thing as this. 11:31.860 --> 11:32.850 Right? 11:32.850 --> 11:34.380 So you're getting the tools by the name, 11:34.380 --> 11:35.910 you're deciding what tools to call. 11:35.910 --> 11:38.730 You are then appending the results into the messages, 11:38.730 --> 11:40.119 as tool messages. 11:40.119 --> 11:44.130 And we've also got this tools condition 11:44.130 --> 11:45.870 and this tools condition, you'll see, 11:45.870 --> 11:47.280 if you look at the literal of that, 11:47.280 --> 11:50.820 we'll just either route to tools or it will route to end, 11:50.820 --> 11:53.850 based on whether there are some tools to to see. 11:53.850 --> 11:55.440 So, it's basically that bit there, 11:55.440 --> 11:57.240 You've got that tools condition. 11:57.240 --> 12:00.930 So, then we've got our chat bots, we add our chat bot node, 12:00.930 --> 12:05.370 we are setting up the graph and we have a tool node. 12:05.370 --> 12:08.610 And notice how the tool node now is much slower, 12:08.610 --> 12:12.720 sorry, much of reduced amount of sort of code, 12:12.720 --> 12:15.000 and we add that to the tools node. 12:15.000 --> 12:17.580 Then we add our conditional edges on the chat bot 12:17.580 --> 12:19.980 where it's over routing directly to tools 12:19.980 --> 12:21.333 or it's routing to end. 12:22.304 --> 12:24.810 We add a edge from the tools back 12:24.810 --> 12:27.540 to the chat bot just in case tools are called. 12:27.540 --> 12:30.093 And we also set the entry 0.2 chat bots. 12:31.410 --> 12:32.550 Right. 12:32.550 --> 12:36.450 So, I'm gonna just get rid of this user session 12:36.450 --> 12:37.800 to which is happening above. (keyboard keys clicking) 12:37.800 --> 12:39.510 So I'm just gonna run these. 12:39.510 --> 12:42.150 And, now, instead of just compiling the graph, 12:42.150 --> 12:44.100 we also add a check pointer here. 12:44.100 --> 12:46.110 And that check pointer will basically, 12:46.110 --> 12:48.480 it's an optional check pointer which will save that 12:48.480 --> 12:51.930 to either an online database or local memory. 12:51.930 --> 12:52.763 If we have a look here, 12:52.763 --> 12:54.390 we haven't changed anything inside 12:54.390 --> 12:56.370 of this LangGraph state machine. 12:56.370 --> 12:58.230 We still have the start node, 12:58.230 --> 13:00.810 which is basically chat bot, 13:00.810 --> 13:02.160 and then we go into tools. 13:02.160 --> 13:04.740 If we have any more tools, we then call those 13:04.740 --> 13:07.710 and add into them until we're finished and then we end. 13:07.710 --> 13:10.350 And if there aren't any tools then we directly end. 13:10.350 --> 13:12.930 Now, we set up a runnable config, 13:12.930 --> 13:14.970 and this is the bit that gives you the ability 13:14.970 --> 13:17.010 to load previous checkpoints 13:17.010 --> 13:19.110 and also update previous checkpoints. 13:19.110 --> 13:21.750 So, it's very important that you have a thread ID 13:21.750 --> 13:25.110 inside of your runnable configuration as you pass 13:25.110 --> 13:28.530 that into the app.stream or the graph.stream. 13:28.530 --> 13:31.200 So you can see here, "Hi, my name is James Phoenix," 13:31.200 --> 13:34.830 we can pass in a message and we pass in the config 13:34.830 --> 13:37.710 and the stream mode just to have a look at the values. 13:37.710 --> 13:40.110 And then we're pretty printing that so you can see 13:40.110 --> 13:42.120 that we've passed in this message 13:42.120 --> 13:44.887 and we then also have, you know, 13:44.887 --> 13:47.160 "How can I help you today?" 13:47.160 --> 13:49.560 And then if we look at a different, 13:49.560 --> 13:51.990 if we look at, like, the same config, 13:51.990 --> 13:54.210 so using the same config as before. 13:54.210 --> 13:55.980 So if we're just having a look at that, 13:55.980 --> 13:57.960 which is the thread ID of one, 13:57.960 --> 14:00.630 then what we can do is say, "Remember my name," 14:00.630 --> 14:02.760 and you can see that it does know 14:02.760 --> 14:05.550 that my name is James Phoenix, which is a good thing. 14:05.550 --> 14:07.470 Now, the problem that we run into 14:07.470 --> 14:09.990 is if we change this thread ID, 14:09.990 --> 14:12.450 you're gonna see the previous message. 14:12.450 --> 14:14.460 It doesn't really know who I am anymore 14:14.460 --> 14:17.190 because it's not using that thread ID, right? 14:17.190 --> 14:21.060 So, the thread ID is the way that the graph is able 14:21.060 --> 14:24.810 to determine specifically what, you know, who you are. 14:24.810 --> 14:29.550 And that's coming from earlier when we loaded the graph, 14:29.550 --> 14:32.700 we actually used a check pointer, right? 14:32.700 --> 14:35.790 So in real time that check pointer is being loaded 14:35.790 --> 14:39.750 and it's being differently attributed to the thread IDs 14:39.750 --> 14:42.300 inside of the SQL lite database. 14:42.300 --> 14:47.300 Now, if we're having a look, we can also modify, 14:47.370 --> 14:50.550 yeah, so the only change that we've made to the thread ID, 14:50.550 --> 14:52.860 but we can also have a look at a few checkpoints 14:52.860 --> 14:53.880 across these different threads. 14:53.880 --> 14:56.220 So, we can get the state of a graph 14:56.220 --> 14:57.660 by taking the configuration. 14:57.660 --> 14:59.790 So, we've got the thread idea of one 14:59.790 --> 15:01.470 and you can see this is the snapshot. 15:01.470 --> 15:03.540 So you get access to things like the DOT values, 15:03.540 --> 15:05.280 which is all the messages. 15:05.280 --> 15:07.230 And there's also a lot more things 15:07.230 --> 15:09.000 that you can have a look at inside of there. 15:09.000 --> 15:10.800 There's, for example, the.next, 15:10.800 --> 15:12.600 which we'll just have a look at over here. 15:12.600 --> 15:15.840 So the snapshot.next shows that there's nothing else 15:15.840 --> 15:18.630 to do on on the graph, right? 15:18.630 --> 15:20.040 But there's lots of different ones, as well. 15:20.040 --> 15:22.230 You can look at the parent configuration, 15:22.230 --> 15:24.270 there's lots of different attributes 15:24.270 --> 15:26.070 that you might find interesting. 15:26.070 --> 15:27.930 There's metadata, for example, 15:27.930 --> 15:30.300 that you can attach specifically to the graph. 15:30.300 --> 15:33.330 And the.next is quite useful for determining 15:33.330 --> 15:35.950 what nodes would be useful to run next 15:36.870 --> 15:38.670 inside of the state machine. 15:38.670 --> 15:41.790 Now, the next node, you know, obviously, has reached an end. 15:41.790 --> 15:43.350 So, next is empty, 15:43.350 --> 15:46.920 but that's basically a little bit on the persistence 15:46.920 --> 15:49.260 and tool usage inside of LangGraph. 15:49.260 --> 15:51.150 Next, we're gonna have a look at 15:51.150 --> 15:53.400 how we can add human in a loop. 15:53.400 --> 15:55.110 So, we've got a chat bot now, 15:55.110 --> 15:57.510 we've added on the Tavily search tool, 15:57.510 --> 15:59.940 and we've also talked around persistence 15:59.940 --> 16:01.800 and how you can use persistence 16:01.800 --> 16:06.240 to specifically remember previous conversations you've had 16:06.240 --> 16:09.180 using a check pointer with the thread ID 16:09.180 --> 16:10.353 inside of LangGraph.