WEBVTT 00:00.080 --> 00:01.240 Hey there Eden here. 00:01.240 --> 00:06.080 And in this video we're going to be implementing the final part of our section. 00:06.240 --> 00:10.120 So we're going to be implementing a version of Adaptive Rack. 00:10.360 --> 00:14.280 And Adaptive Rack is based on this research paper you see right here. 00:14.640 --> 00:23.040 And to be honest, it's a fancy word of simply using a question router to route our question to different 00:23.040 --> 00:23.960 rack flows. 00:24.240 --> 00:27.400 So in this video we're going to be using two rack flows. 00:27.400 --> 00:33.000 So the first one is going to be taking the route to search on the internet and then to continue downstream 00:33.040 --> 00:34.800 on the same logic we have before. 00:35.120 --> 00:40.280 And the second round is going to be to use the retrieval augmentation from our vector store. 00:40.800 --> 00:46.120 So we're first going to take the user's question, and we're going to decide whether the information 00:46.120 --> 00:49.040 is stored in the vector store to answer that question. 00:49.440 --> 00:54.320 And if it's not we're simply going to take the route to web search and answer from there. 00:54.320 --> 01:00.000 And the main thing we're going to do is to implement a question router chain, which will take the question 01:00.000 --> 01:05.040 and decide whether we're going to route it to the web search or to the retrieve node. 01:05.080 --> 01:11.920 We're going to be writing tests to this chain, and we're going to integrate it into our graph by using 01:11.960 --> 01:14.240 the set conditional entry point. 01:15.240 --> 01:16.000 All right. 01:16.000 --> 01:17.280 Let's go to the code. 01:20.760 --> 01:24.360 So let's go and create a new file under chains. 01:24.680 --> 01:27.520 And we want to call this file router. 01:29.000 --> 01:32.600 And this file will hold the implementation of our router chain. 01:33.000 --> 01:35.160 So we'll start from the imports. 01:35.160 --> 01:37.760 And let's import from typing literal. 01:38.120 --> 01:44.400 And if you're not familiar with the literal type like I was a couple of months ago, then it provides 01:44.400 --> 01:50.480 a way to specify that a variable can only take one of predefined set of values. 01:50.600 --> 01:54.320 So this is very useful for validation and type checking. 01:55.000 --> 01:55.640 All right. 01:55.640 --> 02:01.760 We want to import chat prompt template as well because we're going to use structured outputs. 02:01.760 --> 02:06.480 We want to import base model and field from the lecture library. 02:06.920 --> 02:10.520 And also let's import chat OpenAI for our LLM. 02:10.960 --> 02:15.040 Now let's define the class that we want to structure our output. 02:15.040 --> 02:19.280 So we'll call it root query which will inherit from base model. 02:19.760 --> 02:25.520 It will have only one attribute which is called data source and is going to hold only the values of 02:25.520 --> 02:26.400 vector store. 02:26.600 --> 02:31.920 In case we want to search the answer from the vector store, or it's going to hold web search. 02:31.960 --> 02:37.440 If we don't have the answer in the vector store, and we need to perform some external search with Tavileh. 02:38.120 --> 02:44.200 Now when we initialize the field object, if we put here the ellipsis here, then this means that this 02:44.200 --> 02:48.680 field will be required once we instantiate an object of this class. 02:49.160 --> 02:49.880 All right. 02:49.880 --> 02:52.680 And let's also write the description. 02:52.880 --> 02:55.720 And the description is given a user question. 02:55.720 --> 02:59.000 Choose to route it to web search or to vector store. 02:59.280 --> 03:01.320 So we finished with our Pydantic class. 03:01.320 --> 03:07.160 So now it's time to start writing our prompt and to leverage the structured output from LinkedIn. 03:07.560 --> 03:14.000 So we'll create our LLM And we'll create our LLM with structured output with that class. 03:14.160 --> 03:20.400 So basically we are binding now this pedantic class to be called as a function call like we saw earlier 03:20.400 --> 03:21.240 in this course. 03:21.360 --> 03:25.200 And what's going to be interesting here is the system call. 03:26.480 --> 03:32.440 You are an expert at routing a user question to vector store or web search. 03:35.000 --> 03:42.040 The vector store contains documents related to agents, prompt engineering, and adversarial attacks 03:42.040 --> 03:48.160 because that's what we indexed our vector stored with all those articles we ingested in the first video 03:48.160 --> 03:48.960 of this section. 03:49.960 --> 03:53.000 Use the vector store for questions on those topics. 03:53.200 --> 03:56.000 For everything else, use web search. 03:57.520 --> 04:01.160 All right, let's go now and create the prompt itself. 04:01.280 --> 04:04.400 So we'll use chat prompt template from messages. 04:04.760 --> 04:07.320 And we'll first plug in the system message. 04:07.600 --> 04:14.040 And then we're going to plug in the human message which contains the question that we're going to plug 04:14.040 --> 04:15.960 in once we get the user's question. 04:16.280 --> 04:19.320 And lastly, let's go and let's create the chain. 04:19.320 --> 04:22.760 We'll call it question router which will take the route prompt. 04:22.800 --> 04:26.840 We'll pipe it into the router and that's it. 04:26.880 --> 04:28.840 We finished with our chain. 04:30.000 --> 04:35.000 And I hope you're ready because right now we're going to be writing tests to this chain which is super 04:35.000 --> 04:37.280 important for our software hygiene. 04:38.720 --> 04:43.040 And let's go to the file tests.py. 04:44.000 --> 04:48.800 And we want now to import the router question chain. 04:49.000 --> 04:52.480 So let's simply write that string and auto import it. 04:53.560 --> 04:54.600 And let's go up. 04:54.600 --> 04:57.760 And let's just move it after we load the dot env. 04:57.800 --> 04:59.880 So we won't have any issues. 05:00.480 --> 05:01.040 All right. 05:01.040 --> 05:05.800 So now we want to import the um the route query class. 05:06.200 --> 05:07.680 And let's go up. 05:07.680 --> 05:09.480 And we can see we imported it. 05:10.240 --> 05:12.480 And now it's time to write our tests. 05:13.200 --> 05:18.880 So we're simply want to test that we're able to route it to once in the vector store and once to the 05:18.880 --> 05:19.560 web search. 05:19.920 --> 05:27.240 So the first test is going to be dev test raster to vector store, which receives nothing and returns 05:27.240 --> 05:27.640 none. 05:28.360 --> 05:31.880 And here let's put the question to be agent memory. 05:32.000 --> 05:33.680 So it should route to it. 05:34.520 --> 05:43.440 And now we want to invoke the router question chain and will end up with a route query object. 05:43.720 --> 05:48.200 So we'll call it rez which is a route query object. 05:49.920 --> 05:59.200 And it's going to be equal to the question router chain after we invoke it with the question to be agent 05:59.200 --> 05:59.840 memory. 06:02.320 --> 06:03.160 Alrighty. 06:03.520 --> 06:10.840 So let's go now and assert that the data source is going to be equals to vector store, because that's 06:10.840 --> 06:12.120 what we wrote in the prompt. 06:12.160 --> 06:14.360 Remember let's run it. 06:20.520 --> 06:22.880 And we'll wait to get a result. 06:22.880 --> 06:24.720 And we can see this test passed. 06:24.960 --> 06:25.400 Cool. 06:25.600 --> 06:31.880 So now I want to copy this and let's go instead to call it Route to Vector Store. 06:31.920 --> 06:34.120 Let's call it Route to Web Search. 06:34.560 --> 06:38.560 And our question instead of Agent Primary let's write how to make pizza. 06:39.240 --> 06:42.760 And this time this should route us to the web search node. 06:43.040 --> 06:44.080 And let's run it. 06:52.920 --> 06:54.400 And we can see it passed. 06:55.000 --> 06:58.560 And just as a sanity check, let's go and run all of our tests. 06:58.760 --> 07:04.960 And I have to admit, like every developer, seeing all of those green check marks is very satisfying. 07:05.600 --> 07:09.520 So let's wait and and see that everything passes. 07:10.040 --> 07:13.840 By the way, we can make optimizations to all of those tests. 07:13.840 --> 07:17.760 For example, we can run them concurrently because they're not dependent on each other. 07:18.040 --> 07:22.280 But um, we're not going to cover it now, maybe later in this course. 07:22.960 --> 07:23.680 All right. 07:23.680 --> 07:28.680 So all of the tests passed and we are pretty confident with our chain. 07:29.320 --> 07:30.120 All right. 07:30.120 --> 07:35.240 So let's have a look on what we're going to be implementing right now in our graph py file. 07:35.560 --> 07:38.120 Eventually this is how our graph should look like. 07:38.320 --> 07:40.320 We should start from the start node. 07:40.560 --> 07:46.680 Then we should have this conditional branch which would either go to retrieve or either go to web search 07:46.800 --> 07:49.960 depending on our router that we just wrote. 07:50.360 --> 07:56.200 And for this we're going to be introducing something new in graph, which is called a conditional entry 07:56.240 --> 08:03.040 point, which basically is a conditional edge with the first node of the entry point. 08:03.320 --> 08:06.040 So this is basically going to be our routing. 08:06.160 --> 08:10.760 And from them we've already implemented all of the rest of the graph. 08:13.000 --> 08:17.080 All right so let's go back to the graph.py file. 08:17.400 --> 08:24.200 And now we want to import the router and the router query chain from the router.py file. 08:24.360 --> 08:26.880 So this is what happens in line seven. 08:27.640 --> 08:28.160 Cool. 08:28.200 --> 08:35.160 So let's now implement the root question function, which is going to receive the state and return us 08:35.200 --> 08:38.600 a string which is going to be the next node to execute. 08:38.600 --> 08:41.080 And you can see the illustration on the right. 08:42.040 --> 08:42.640 All right. 08:42.640 --> 08:47.160 Let's add a print that we are in the root question. 08:47.680 --> 08:51.280 And let's fish out the question from our graph state. 08:51.560 --> 08:56.240 Now we'll simply run the question router chain with the invoke method. 08:56.240 --> 09:00.960 Plugging in the user's question which we fished out from the state. 09:01.520 --> 09:07.840 And the result we're going to save into a source variable, which is going to be of the type root query. 09:07.920 --> 09:14.720 Because I remind you that the question router chain returns this kind of object because we use the structured 09:14.720 --> 09:15.560 output chain. 09:16.840 --> 09:17.360 All right. 09:17.360 --> 09:19.640 So now we have the result. 09:19.680 --> 09:25.800 Now if the data source is web search we want to route it into the web search node. 09:25.800 --> 09:28.320 So we'll simply return web search. 09:28.600 --> 09:35.720 And if the answer is vector store, then we simply want to route it into the retrieve node to perform 09:35.720 --> 09:36.320 the retrieval. 09:36.320 --> 09:37.160 Augmentation. 09:39.520 --> 09:40.120 All right. 09:40.120 --> 09:42.680 So let's add the conditional entry point. 09:42.880 --> 09:47.840 And for that we want to use the function set conditional entry point. 09:49.600 --> 09:55.280 And now we want to add our route function that we just implemented. 09:56.160 --> 09:58.880 And let's add the path map. 09:58.880 --> 10:01.600 So web search is going to be mapped to web search. 10:01.600 --> 10:04.080 And retrieve is going to be mapped to retrieve. 10:04.360 --> 10:05.080 All right. 10:05.080 --> 10:07.320 So that's it for the implementation. 10:07.920 --> 10:08.600 Let's go. 10:08.600 --> 10:10.240 And let's now run the code. 10:10.240 --> 10:11.960 So I'm going to go to the main file. 10:12.120 --> 10:14.920 And let's run it with what is agent memory. 10:14.920 --> 10:19.160 So our original prompt from the earlier in this section. 10:20.200 --> 10:25.720 So we can see that we are now starting to first route the question. 10:26.600 --> 10:32.520 And now we decide to route it into the vector store and to perform the retrieval, augmentation. 10:32.720 --> 10:36.400 And from then the flow remains, as we know already. 10:37.480 --> 10:40.440 And let's wait for the final answer. 10:40.600 --> 10:43.080 And we indeed get the answer that we want. 10:43.120 --> 10:50.080 We can go to Blanc-smith, and we can check out all of the traces and all of the code that's running 10:50.680 --> 10:51.120 now. 10:51.160 --> 10:53.800 Note that we didn't add any other node here. 10:53.800 --> 10:58.400 We simply added a conditional edge from our entry point. 10:59.240 --> 11:00.640 And let's go to the code. 11:00.640 --> 11:04.600 And let's change the query to how to make pizza. 11:06.360 --> 11:07.680 And let's run it again. 11:07.720 --> 11:14.160 So now we're supposed to root into web search and to start running traveling and to continue from there. 11:18.080 --> 11:20.560 So you can see that we routed into web search. 11:20.560 --> 11:21.080 Indeed. 11:21.320 --> 11:24.360 And yeah, that's pretty much it. 11:24.400 --> 11:26.720 And we got here the answer. 11:27.880 --> 11:28.840 Alrighty. 11:28.840 --> 11:33.280 So we are pretty much done with our advanced drag flow.