WEBVTT 00:00.160 --> 00:00.920 All right. 00:00.920 --> 00:04.280 So let's go and implement our retrieval pipeline. 00:04.880 --> 00:08.280 So I'm going to start by creating a main.py file. 00:08.280 --> 00:11.720 And here is going to be the retrieval part of our project here. 00:11.760 --> 00:18.960 So let's go now and start with the imports and the initialization of the object that we'll be using. 00:18.960 --> 00:20.360 And let's start. 00:20.560 --> 00:27.120 So first I want to import the OS module because we're going to be accessing environment variables. 00:27.600 --> 00:33.160 And now let's go and import load env because we're going to be loading all the environment variables 00:33.160 --> 00:35.120 from the env file over here. 00:35.720 --> 00:39.400 Let's go now and import from the chat prompt template. 00:39.400 --> 00:43.440 Because this is what's going to form the react prompt we're going to be using. 00:44.200 --> 00:46.640 We want to import human message. 00:46.640 --> 00:49.400 This is what is going to be invoking the pipeline. 00:50.160 --> 00:58.480 And let's go and import from OpenAI the chat OpenAI for our ML model and OpenAI embeddings for our embeddings 00:58.480 --> 01:01.290 modules, which we covered in the previous section. 01:01.290 --> 01:05.970 And also let's go and import like previously explained, the pine cone vector store. 01:06.170 --> 01:10.050 And let's start with the initialization of all of the objects. 01:10.330 --> 01:13.450 So first we want to load all the environment variables. 01:13.810 --> 01:17.410 And let's now go and print initializing components. 01:17.410 --> 01:20.930 So now we're going to be initializing every one of those objects here. 01:21.210 --> 01:23.970 So let me go now and run everything as a sanity check. 01:24.770 --> 01:27.570 And we can see we are printing this indeed. 01:27.930 --> 01:28.410 Cool. 01:28.890 --> 01:33.010 So let's start with initializing OpenAI embeddings. 01:33.650 --> 01:35.530 And let's also initialize the LLM. 01:35.530 --> 01:39.010 And here we're using the link chain OpenAI defaults. 01:39.010 --> 01:41.250 And we'll see which LLM is going to be using. 01:42.610 --> 01:46.370 Let's go now and initialize our vector store object. 01:46.370 --> 01:52.450 So in order to initialize it we need to give it the index name which we created earlier in the videos. 01:52.690 --> 01:58.490 And we need to give it the embeddings module, very similar to what we did in the previous video. 01:58.930 --> 02:04.370 And now we want to use this vector store and we want to use its searching capabilities. 02:04.370 --> 02:09.770 So we want to be able to make similarity searches to find these relevant contexts in direct pipeline. 02:10.130 --> 02:16.130 So if earlier let's go to the ingestion.py file we use the from document method. 02:16.410 --> 02:18.850 Let's go back now to main.py. 02:19.450 --> 02:24.010 Now we want to go and we want to take this vector store. 02:24.050 --> 02:27.570 And it has a method which is called as retriever. 02:27.970 --> 02:30.170 And this function is going to return us. 02:30.170 --> 02:32.050 Let's go and open the docs here. 02:32.050 --> 02:36.130 It's going to return us an object of vector store retriever. 02:36.330 --> 02:43.250 So this vector store retriever from link chain I'm not going to go too deep into this in this section 02:43.250 --> 02:49.570 right now, but it's going to give us a vector store with searching capabilities that the vendors of 02:49.570 --> 02:51.130 the vector store implemented. 02:51.330 --> 02:57.690 And eventually these vector store retriever class is going to have this searching capability. 02:57.730 --> 03:01.220 It's going to have a search function, which we're going to be using now. 03:01.500 --> 03:05.380 So when we initialize it we're going to give it search queries. 03:05.420 --> 03:06.660 K equals three. 03:06.820 --> 03:12.660 So this means every time we want to search in the vector store and get those relevant chunks those relevant 03:12.660 --> 03:13.420 contexts. 03:13.620 --> 03:19.740 I want to be limiting it to only three top documents that are going to be used. 03:19.940 --> 03:27.020 So if there are maybe ten relevant documents, it's going to sort them by order of importance of what's 03:27.020 --> 03:29.620 the most relevant and what's the least relevant. 03:29.620 --> 03:31.940 And I want to take the top three. 03:32.780 --> 03:37.020 And let me now go and initialize the prompt template here. 03:37.380 --> 03:41.020 So I'm going to initialize it with a very simple prompt. 03:41.020 --> 03:43.180 But it's a very powerful prompt. 03:43.500 --> 03:44.940 The prompt goes as follows. 03:44.980 --> 03:48.780 Answer the question based only on the following context. 03:49.020 --> 03:51.460 Here we have the placeholder for the context. 03:51.700 --> 03:57.460 Here we have the original question of the user and provide a detailed answer. 03:57.660 --> 03:59.460 So this context here. 03:59.460 --> 04:03.100 This is going to be the augmentation part of our prompt. 04:03.100 --> 04:06.100 So the question here is going to be our original prompt. 04:06.100 --> 04:08.220 And this is going to be the augmentation. 04:08.220 --> 04:11.780 So from the retrieval augmentation generation of Rag. 04:13.300 --> 04:18.020 And now let's go and implement a very simple auxiliary function. 04:18.020 --> 04:23.940 So we want to implement format docs which is going to receive docs which are going to be linked in documents. 04:24.380 --> 04:31.860 And this function is going to take the documents that were retrieved from the vector store, and it's 04:31.860 --> 04:34.780 simply going to format them nicely into a nice string. 04:34.940 --> 04:36.500 So you can see all that it's doing. 04:36.500 --> 04:39.660 It's going to iterate through all the documents. 04:39.700 --> 04:41.180 This is the for loop over here. 04:41.580 --> 04:45.420 Every document we saw before has a page content. 04:45.580 --> 04:48.620 So we're going to take this page content which is a string. 04:48.620 --> 04:51.900 And we are simply going to append newlines before it. 04:52.140 --> 04:57.460 So if we have a bunch of documents we're simply going to format them as one string. 04:57.780 --> 05:02.790 And this one string is eventually what's going to be sent here to this context, right? 05:03.590 --> 05:09.670 So those are all the initialization of the objects, and those are all the building blocks that we're 05:09.670 --> 05:10.830 going to be using now. 05:11.750 --> 05:12.230 Cool. 05:12.270 --> 05:14.910 So now let's go try to run everything. 05:14.910 --> 05:18.390 Let's see that everything is going to be initialized. 05:18.430 --> 05:18.790 Cool. 05:18.830 --> 05:19.950 We didn't get any error. 05:19.990 --> 05:21.550 This was a sanity check. 05:22.910 --> 05:23.350 Cool. 05:23.550 --> 05:27.430 So now let's go and use our main runner here. 05:27.670 --> 05:31.750 Now I wrote if name equals main, we're going to be print retrieving. 05:31.750 --> 05:35.790 And the query is going to be what is pinecone in machine learning. 05:35.950 --> 05:40.550 So first let me go and show you what we're going to get if we're not going to be using rag. 05:40.790 --> 05:45.350 So here I pasted a raw invocation which is simply going to use the LM. 05:45.470 --> 05:50.830 It's going to send it in the human message, the content of the query, which is what is pinecone in 05:50.830 --> 05:51.550 machine learning. 05:51.790 --> 05:53.310 And then it's going to print it. 05:53.510 --> 05:56.910 So let's go and let me show you what we're going to be getting. 05:57.110 --> 05:58.840 So this is a raw invocation. 05:58.840 --> 06:00.120 We are not using rag. 06:00.560 --> 06:02.160 So here we can see the answer. 06:02.360 --> 06:09.080 A pinecone algorithm is a method used in machine learning to search for the best configuration hyperparameters. 06:09.600 --> 06:15.400 This is not the answer that we want, because we want to talk about pinecone as the vector store. 06:15.800 --> 06:16.240 All right. 06:16.240 --> 06:17.760 And let me even go to links. 06:18.520 --> 06:22.400 Let's go to our links project here that I initialized. 06:22.400 --> 06:23.360 Here we have the chain. 06:23.360 --> 06:25.280 What is pinecone in machine learning. 06:25.720 --> 06:27.080 And here we can see the answer. 06:27.080 --> 06:31.160 By the way the model that we're using is GPT 3.5 turbo. 06:31.440 --> 06:33.920 So just for the sake of it let's go and change here the model. 06:33.920 --> 06:37.680 Let's try to use model GPT 5.2. 06:51.920 --> 06:53.240 Now we can see the answer. 06:53.240 --> 06:58.320 Python is a managed vector store database using machine learning to store index. 06:58.360 --> 06:58.640 Okay. 06:58.680 --> 07:00.280 So this is a nice answer. 07:00.280 --> 07:02.240 So this is like the answer we wanted. 07:02.240 --> 07:07.640 And by the way the difference here and this is by the way one of the motivations of using Rag because 07:07.680 --> 07:15.480 GPT 5.2 was trained in 2025 and pinecone was much established by then. 07:15.480 --> 07:19.640 So we had lots of documents and lots of training data on pinecone as the vector store. 07:19.800 --> 07:23.600 And before that it was not so GPT 3.5. 07:23.640 --> 07:26.240 As we saw before, it hallucinated the answer. 07:26.280 --> 07:28.120 This is also something interesting to see. 07:28.160 --> 07:28.560 Cool. 07:28.560 --> 07:31.760 So now let's go and use a chat. 07:31.760 --> 07:34.320 OpenAI with GPT 3.5. 07:34.720 --> 07:38.640 And we saw what was the answer without using Rag. 07:38.680 --> 07:43.160 And now let's go and implement the Rag pipeline with all the building blocks. 07:43.200 --> 07:43.760 All right. 07:43.760 --> 07:48.440 So now let's go and implement the retrieval part of our pipeline. 07:48.600 --> 07:54.800 So first let me define a function which is called retrieval chain without link chain expression language. 07:55.160 --> 08:01.450 And this is a function which is going to receive a string, and it's going to return us a response of 08:01.450 --> 08:02.090 the LM. 08:02.130 --> 08:07.770 The way we're going to be implementing now is without using linked chain expression language, and this 08:07.770 --> 08:12.130 is simply to show you the flow and to show you what's happening under the hood. 08:12.330 --> 08:16.850 And this is to show you how we can do it without link chain and how it's going to be looking like. 08:17.090 --> 08:19.210 So let's go and start and implement this. 08:19.210 --> 08:21.410 And this is the most naive implementation. 08:21.730 --> 08:27.570 After we're going to be doing that, we are going to implement it with linked chain expression language, 08:27.810 --> 08:33.010 which is going to yield us with a much more elegant solution and much more capabilities. 08:33.250 --> 08:37.170 So here in the description this is going to be a simple retrieval chain. 08:37.210 --> 08:38.610 It's not going to be really a chain. 08:38.650 --> 08:41.650 It's simply going to be a bunch of function invocations. 08:41.970 --> 08:45.010 And here we're going to manually retrieve the documents. 08:45.010 --> 08:46.970 We're going to manually format them. 08:46.970 --> 08:49.370 And we're going to generate the answer. 08:49.570 --> 08:54.170 So in this implementation we have a bunch of limitations which are going to be solved. 08:54.490 --> 08:58.380 In this video we are going to invoke everything manually. 08:58.660 --> 09:00.820 We don't have any streaming support. 09:00.820 --> 09:02.700 We don't have any async support. 09:02.740 --> 09:06.380 It's hard to take these and compose it in other chains. 09:06.380 --> 09:10.020 And this is going to be error prone and it's going to be harder to maintain. 09:10.020 --> 09:14.460 And we're going to see how we're going to be solving every one of those limitations in the length chain 09:14.460 --> 09:18.300 expression language implementation, which is going to be in this video. 09:18.460 --> 09:24.420 So the first thing we want to do is we want to take the user query the question, what is pinecone in 09:24.420 --> 09:25.180 machine learning. 09:25.340 --> 09:28.580 And we want to get the most relevant documents. 09:28.900 --> 09:32.500 We want to perform similarity search in the pinecone vector store. 09:32.740 --> 09:36.100 So when we're going to be using our retriever dot invoke. 09:36.380 --> 09:41.220 So retriever right over here is actually a link chain runnable. 09:41.500 --> 09:48.900 So if I'm going to actually go over here and this is going to return us a vector store retriever object. 09:48.980 --> 09:53.500 This object is inheriting from base retriever and base retriever. 09:53.500 --> 09:55.100 Here is a runnable. 09:55.500 --> 09:58.660 So this is going to have the invoke method here. 09:58.700 --> 10:05.580 Now when we're going to be using a vector to retrieve our in length chain because it has an invoke method, 10:05.580 --> 10:09.260 its invoke method, it's going to run something. 10:09.260 --> 10:14.820 It's going to run a bunch of things, but eventually it's going to run this function, get relevant 10:14.820 --> 10:21.140 documents, which is going to take our query, and it's going to return us a list of documents which 10:21.140 --> 10:23.540 are going to be the most relevant documents. 10:23.780 --> 10:25.780 So these get relevant documents. 10:25.780 --> 10:28.820 This is eventually going to be implemented by the vendor. 10:28.940 --> 10:31.380 Pinecone have their own way of doing things. 10:31.380 --> 10:33.260 It's going to use the pinecone SDK. 10:33.300 --> 10:37.900 If we're going to be using chroma it's going to be using another implementation. 10:37.900 --> 10:40.500 And this is eventually what's going to be running. 10:40.620 --> 10:44.380 And we have also an async version of getting the relevant documents. 10:44.380 --> 10:50.660 But eventually when we are going to be running this invoke method, it's going to be going and doing 10:50.660 --> 10:50.940 this. 10:50.940 --> 10:53.940 And we're going to see this in the length trace by the way. 10:53.940 --> 10:55.260 And let's now assume. 10:55.260 --> 11:01.590 So this is going to magically give us the most relevant documents from our original query here. 11:01.590 --> 11:05.070 And because we initialized it with k equals three. 11:05.390 --> 11:10.030 So this means that this is going to be a list of three link chain documents. 11:10.270 --> 11:12.750 So we have now the relevant documents. 11:12.950 --> 11:17.390 Next we want to go and let's format them into a string. 11:17.670 --> 11:20.110 So we're going to be using the format docs. 11:20.110 --> 11:24.510 And it's going to give us here the context which is going to be a string. 11:24.550 --> 11:24.910 Cool. 11:24.910 --> 11:29.230 So right now we have the question we have the context for this question. 11:29.390 --> 11:32.710 So now we want to go and we want to take this prompt. 11:32.710 --> 11:37.630 And we simply want to plug the value of the context and the value of the questions here. 11:37.830 --> 11:40.270 And we want to go and send this to the LLM. 11:40.310 --> 11:43.350 So now we are taking the prompt template. 11:43.390 --> 11:45.550 We are going to use the format messages. 11:45.550 --> 11:49.230 And here we're going to provide it with the context equals context. 11:49.230 --> 11:50.990 And the questions equals query. 11:51.390 --> 11:54.710 So question here is matching. 11:54.710 --> 12:00.470 Now this placeholder here in the query is the value that we're going to be giving it. 12:00.790 --> 12:07.590 So now this is going to give us a list of messages holding one message, which is going to be everything 12:07.590 --> 12:09.110 we want to send to the LM. 12:09.350 --> 12:16.030 And now let's go and invoke our LM with the list of messages, which is going to be one message. 12:16.190 --> 12:20.150 And we simply want to return the content of that response here. 12:20.270 --> 12:26.030 So this is the entire implementation the manual implementation of the Reg pipeline over here. 12:26.230 --> 12:26.750 All right. 12:26.750 --> 12:33.470 And now let me paste in the snippet which is simply going to print that we want to invoke the implementation 12:33.470 --> 12:35.350 without an expression language. 12:35.590 --> 12:37.870 We're going to be taking this function. 12:38.030 --> 12:41.710 And we're going to be running it with the user's query. 12:42.070 --> 12:43.870 And we're going to be printing. 12:43.870 --> 12:46.430 Now the result nothing special over here. 12:46.430 --> 12:46.990 All right. 12:47.030 --> 12:47.710 Let me now. 12:47.750 --> 12:50.350 So let's go now and run everything. 12:54.720 --> 12:59.480 So this is the raw invocation here we see we get the wrong answer. 12:59.480 --> 13:02.080 And let's see the question answered with rag. 13:02.320 --> 13:08.040 And here we can see pinecone is a fully managed cloud based vector database specifically designed for 13:08.040 --> 13:14.200 businesses and organizations looking to build and deploy large scale machine learning applications. 13:14.520 --> 13:18.160 So this is the answer that we wanted working with Rag here. 13:18.160 --> 13:20.160 So now I want to show you in debug. 13:20.280 --> 13:22.640 And let's see step by step what's happening here. 13:22.920 --> 13:24.880 So let me go to the function here. 13:25.160 --> 13:27.080 And let me now put a breakpoint. 13:30.200 --> 13:31.720 Let's go run it in debug. 13:36.120 --> 13:36.680 All right. 13:36.680 --> 13:41.160 So here we can see we have now the query what is pinecone machine learning. 13:42.000 --> 13:44.080 Let me now go and execute this. 13:50.280 --> 13:51.600 So here we can see. 13:51.640 --> 13:54.360 Now we have documents variable. 13:54.610 --> 13:58.690 and it's going to contain a list of three link chain documents. 13:59.090 --> 14:06.330 And every document here is going to have its page content attribute here, which is going to be the 14:06.370 --> 14:08.010 data of this document here. 14:08.010 --> 14:13.210 So these are the most relevant documents that pinecone found for us. 14:13.410 --> 14:19.650 Now we want to take those documents and now we want to format them as strings. 14:19.850 --> 14:23.250 So let's go and run this format docs function. 14:24.090 --> 14:26.530 And now we have the context variable. 14:26.530 --> 14:29.250 And now we can see this is a string okay. 14:29.290 --> 14:32.290 So now let me copy here this expression. 14:32.290 --> 14:34.090 Let me open it here and let me paste it. 14:34.090 --> 14:38.490 And we can see it's a very big string of all the documents combined. 14:38.730 --> 14:44.770 So now we want to go and we want to initialize the prompt template with the context of the documents. 14:44.770 --> 14:46.370 Here we can see all the documents. 14:46.730 --> 14:49.490 And we want now to give it the query. 14:49.490 --> 14:52.010 And we want to initialize it with the query. 14:52.050 --> 14:54.490 What is pinecone in machine learning. 14:54.730 --> 14:55.930 Let's go and do that. 14:56.490 --> 14:58.370 Let's go and look for messages. 14:58.370 --> 15:02.490 So it gave us a messages list with one message. 15:02.810 --> 15:06.450 And each content is the prompt we had before. 15:06.570 --> 15:09.730 Answer the questions based on the following context. 15:09.730 --> 15:13.210 And then it pasted the context and gave the question here. 15:13.490 --> 15:14.090 Cool. 15:14.090 --> 15:17.810 And let's go now and invoke now the LM. 15:17.970 --> 15:21.130 And this is a regular LM invocation right now. 15:21.570 --> 15:23.890 And we get here a response. 15:23.890 --> 15:25.650 Let's go and check out this response. 15:25.650 --> 15:30.690 The response is an AI message and its content is the answer. 15:30.850 --> 15:34.090 So now I want to show you the Lang Smith trace. 15:34.570 --> 15:40.050 And because here in these functions we did not really use a Lang chain chain. 15:40.210 --> 15:47.090 And we did not use the chain expression language, then we simply used all linked chains components 15:47.130 --> 15:48.010 separately. 15:48.290 --> 15:56.020 So all the traces are not going to be organized under one component of a chain, and this is not convenient 15:56.020 --> 15:56.420 at all. 15:56.460 --> 15:59.620 Okay, so now let's go and review the traces here. 15:59.620 --> 16:03.500 So let's take the first step of retrieving the relevant documents. 16:03.900 --> 16:08.900 So here we can see we have a vector store retrieval right over here. 16:09.300 --> 16:12.780 And this is the run of this line over here. 16:13.620 --> 16:17.660 So here we can see the input is what is pinecone in machine learning. 16:17.900 --> 16:24.980 When we ran it in evoke, the output we get was those three documents of the data that are the most 16:24.980 --> 16:26.180 relevant documents. 16:26.460 --> 16:33.180 So this was the similarity search that pinecone did for us under the hood here and here we can see we 16:33.180 --> 16:34.700 have the text. 16:34.700 --> 16:36.540 So this is the text of the document. 16:36.540 --> 16:41.180 And here we even have the source which I did not show you in the debug console. 16:41.180 --> 16:46.020 So now after we did that we went we formatted nicely. 16:46.020 --> 16:48.020 This is nothing to do with Lang Smith. 16:48.020 --> 16:50.780 And here we simply populated the prompt template. 16:50.780 --> 16:54.900 And this is not going to show in Lang Smith as well because it's not under the chain. 16:55.140 --> 17:01.380 And finally, we invoke the LM with all the relevant data in the augmented prompt. 17:01.380 --> 17:04.780 So we can see the LM invocation right over here. 17:05.060 --> 17:09.300 So the final call to LM is the original prompt. 17:09.460 --> 17:11.140 Here we have the context. 17:11.340 --> 17:14.300 And here we have the question that we plugged in. 17:14.340 --> 17:18.180 And this is the answer that the LM returned us. 17:18.420 --> 17:23.460 So we can see tracing this and debugging this is going to be a bit harder. 17:23.620 --> 17:24.340 All right. 17:24.340 --> 17:30.660 So what we did now in the implementation is implement everything very very naively. 17:30.860 --> 17:33.780 And it has a bunch of disadvantages. 17:33.780 --> 17:38.340 But the main disadvantage is that it's going to be very hard to trace it. 17:38.380 --> 17:44.820 So in the next video we're going to implementing the same functionality but binding everything under 17:44.820 --> 17:48.020 a link chain chain with the link chain expression language. 17:48.020 --> 17:50.820 And this is going to be a much better implementation. 17:50.820 --> 17:53.340 And we're going to see all the reasons why it's better.