WEBVTT 00:00.320 --> 00:01.160 All right. 00:01.160 --> 00:04.720 So finally in this section we're going to build our graph. 00:04.720 --> 00:08.040 And we're going to connect all the nodes and edges together. 00:08.400 --> 00:10.560 So let's go to the code. 00:11.200 --> 00:15.920 We first want to go to the init file under the nodes module. 00:16.240 --> 00:20.000 So here we want to import all of the nodes we created. 00:20.000 --> 00:25.080 So it's going to be the generate node the create documents node the retrieve node the the web search 00:25.080 --> 00:25.520 node. 00:25.720 --> 00:29.360 And we want to be able to import them from outside the package. 00:29.480 --> 00:32.480 So that's why we're going to use the underscore all. 00:32.800 --> 00:36.760 And here we're going to put the string names of those nodes. 00:36.800 --> 00:41.880 So this will make all of those nodes importable from outside packages. 00:42.280 --> 00:43.000 Alrighty. 00:43.040 --> 00:45.120 Now I want to go to const py file. 00:45.120 --> 00:47.160 And let's define a bunch of constants. 00:47.320 --> 00:52.400 And all of those cons are going to be the node names that we're going to refer to in the graph. 00:52.480 --> 00:56.040 And the reason we do this is to avoid code duplication. 00:56.040 --> 01:00.160 So every time we'll reference the nodes we'll reference the const. 01:00.600 --> 01:06.040 And if we want to change the name from some reason we will only need to change it in one place. 01:06.280 --> 01:08.640 So that's why I put it in the const file. 01:08.680 --> 01:12.200 So we defined the retrieve const with capital letters. 01:12.200 --> 01:17.480 And the value is going to be lowercase retrieve which is going to be the name of the retrieve node. 01:17.760 --> 01:21.440 Similarly to what we see on the graph of our right. 01:23.400 --> 01:32.440 Alrighty, now it's time to go to the graph.py file, and let's finally build an orchestrator graph. 01:33.080 --> 01:35.680 So we first want to start with the imports. 01:36.040 --> 01:39.880 Let's start by importing load dot env to load the environment variables. 01:40.320 --> 01:45.040 We also want to import from graph the add node and the state graph. 01:45.520 --> 01:52.920 And we want to import all of our costs that we defined earlier which are going to be the nodes names. 01:53.600 --> 01:58.520 And now we want to import all of the nodes we created in the earlier sections. 01:58.880 --> 02:05.440 So we'll import the generate node, the greet documents, the retrieve node and finally the web search 02:05.440 --> 02:05.920 node. 02:06.480 --> 02:11.970 And finally we want to import the graph state which we created in the states py paid for. 02:13.730 --> 02:20.130 All right, so if we take a look at the graph from our right, then we can see that after the great 02:20.130 --> 02:22.330 documents node we have two edges. 02:22.330 --> 02:23.850 One is going to the web search. 02:23.850 --> 02:28.690 And this is if we found at least one document that is not relevant to the question. 02:29.050 --> 02:31.730 And we have an edge to the generate node. 02:31.770 --> 02:37.210 So this is if all documents are valid and are related to the question. 02:37.410 --> 02:41.010 And this is going to be our graph conditional edge. 02:41.370 --> 02:47.730 So let's go and define the function that will decide which node are we going next. 02:47.850 --> 02:53.930 So it's going to output either the web search node or the generate node according to the conditions 02:53.930 --> 02:55.250 we described just now. 02:55.850 --> 03:01.450 So we'll define the function we'll call it decide to generate which will accept the graph state. 03:01.890 --> 03:06.890 And if the web search flag in our state is going to be true. 03:07.210 --> 03:12.330 And this means that we found a document that is not relevant to the user's query. 03:12.530 --> 03:16.570 So this is the heuristic that in this case we want to search online. 03:16.770 --> 03:19.810 So here we would like to return the web search node. 03:19.850 --> 03:21.810 Now notice I'm returning the const. 03:21.810 --> 03:22.930 And this is the name. 03:23.570 --> 03:29.490 And if not and this means that all of the documents we retrieved were relevant to the query. 03:29.490 --> 03:31.330 So here we want to return generate. 03:31.330 --> 03:34.930 So the graph execution will move to the generate node. 03:35.330 --> 03:40.010 And now let's do the easiest part which is connecting everything together. 03:40.290 --> 03:42.210 So we'll create our flow. 03:42.250 --> 03:43.490 We'll call it workflow. 03:43.490 --> 03:49.290 And it's going to be the graph state graph where the state is the state we defined. 03:49.730 --> 03:55.370 And we'll use the add node functions to add all of the nodes currently not connected. 03:55.570 --> 04:00.570 So we added four nodes and they retrieve node is going to run the retrieve function. 04:00.610 --> 04:04.490 Now notice that the retrieve name is lowercase like we see in the picture. 04:04.490 --> 04:05.410 From our right. 04:05.770 --> 04:11.570 We have the Great Documents node which is going to run the Great Documents function, the generate node, 04:11.570 --> 04:12.770 and the web search node. 04:13.330 --> 04:16.690 Let's now add the entry point to be the retrieve node. 04:16.690 --> 04:20.610 So this is the first function which is going to run to retrieve the documents. 04:20.810 --> 04:24.450 We then want to move into the Great Documents node. 04:24.450 --> 04:30.010 So we'll add an edge between the retrieve node and between the Great Documents node. 04:30.410 --> 04:37.210 And now we want to add the conditional edge from the great nodes to either web search or the generate 04:37.250 --> 04:37.770 node. 04:38.170 --> 04:43.450 So we're going to use the add conditional edge function from the great documents. 04:43.450 --> 04:44.770 This is the source node. 04:45.050 --> 04:48.650 The function which will decide which node is going to be executed. 04:48.690 --> 04:52.170 Next is going to be decide to generate which we implemented. 04:52.450 --> 04:55.410 And the third argument which is going to be the path map. 04:55.410 --> 05:00.450 I'm going to put here a dictionary from web search to web search and from generate to generate. 05:00.930 --> 05:06.850 And the reason I put it here is simply to show you that there is this option, and what this will do 05:06.850 --> 05:09.370 if we put here other names in the values. 05:09.370 --> 05:16.970 For example, if instead of web search as the value would put end, then this would be a simply different 05:16.970 --> 05:18.330 mapping to go to. 05:18.770 --> 05:24.490 So it's useful when the function you implement that decides which node to go to is not returning the 05:24.490 --> 05:25.250 node names. 05:25.810 --> 05:28.330 And again in this function we we implement it? 05:28.370 --> 05:29.530 We don't really need them. 05:29.530 --> 05:32.490 I simply wanted to show you that this option exists. 05:33.850 --> 05:41.290 Alright, so after we execute the web search node, we have the results and we're ready to go to the 05:41.290 --> 05:42.210 generate node. 05:42.210 --> 05:45.570 So let's add the edge from web search to generate. 05:46.330 --> 05:50.650 And now we want to add the last edge from generate to end. 05:51.210 --> 05:52.010 And that's it. 05:52.050 --> 05:53.930 We finished building our graph. 05:54.330 --> 05:55.650 Let's compile it. 05:56.010 --> 05:58.010 And we also want to print it. 05:58.050 --> 05:59.970 We'll print it into graph PNG. 06:01.250 --> 06:05.090 And finally we are done with our graph. 06:05.090 --> 06:08.890 So now let's go to the main.py file. 06:08.890 --> 06:11.410 And let's import the graph. 06:11.410 --> 06:15.490 And let's invoke it with the questions of what is agent memory. 06:16.490 --> 06:19.250 Alrighty, let's go and run this baby. 06:21.450 --> 06:25.290 And from our prints we should see the graph structure. 06:27.210 --> 06:29.250 So we can see we achieved the documents. 06:29.250 --> 06:31.010 We checked relevance. 06:31.050 --> 06:33.330 We saw which documents are relevant or not. 06:33.370 --> 06:39.740 We saw that one document was not relevant, so we executed the external search. 06:40.180 --> 06:44.260 We then went to the generate node and finally we had the response. 06:44.500 --> 06:46.700 And the response we got was pretty good. 06:46.700 --> 06:48.540 Let's go and check out the graph png. 06:48.580 --> 06:51.580 So this is the exact graph we see from our right. 06:52.260 --> 06:57.540 And let's go to lengths to trace our graph execution. 06:58.260 --> 07:00.980 And let's take the last trace over here. 07:01.460 --> 07:06.540 And we can see we have all the nodes that we executed from their order. 07:07.100 --> 07:07.620 We can see. 07:07.660 --> 07:11.300 For example let's take a look at the web search. 07:11.300 --> 07:17.860 And we can see that we executed here with the search item of what is agent memory. 07:20.220 --> 07:21.060 Alrighty. 07:21.100 --> 07:27.060 And last thing, if you want to check out the code, please go and visit the GitHub repository in the 07:27.060 --> 07:32.420 branch line graph and you should see the entire implementation you just saw there. 07:32.740 --> 07:37.380 Updated with the latest line chain and line graph versions.