WEBVTT 00:00.360 --> 00:03.000 Teacher: A lot of LLM applications involve 00:03.000 --> 00:07.260 somewhat connecting LLMs to external data sources. 00:07.260 --> 00:08.310 Like we saw in the course, 00:08.310 --> 00:10.170 where we took Medium articles 00:10.170 --> 00:12.480 and we talked the link chain documentation 00:12.480 --> 00:14.670 and we QAed over them 00:14.670 --> 00:18.780 by using the retrieval augmentation generation method. 00:18.780 --> 00:22.680 But a prerequisite for doing it is to ingest the data 00:22.680 --> 00:27.680 into a format that the LLM can easily understand it. 00:27.750 --> 00:29.790 So, most of the time it'll mean 00:29.790 --> 00:31.530 that we need to take our data 00:31.530 --> 00:33.510 and ingest it into a vector store 00:33.510 --> 00:35.070 like we showed in the course. 00:35.070 --> 00:39.150 But before we do it, we need to first split it 00:39.150 --> 00:41.730 or chunkify it into smaller chunks 00:41.730 --> 00:44.520 because we can't just put all of the data as is 00:44.520 --> 00:45.630 in the vector store. 00:45.630 --> 00:47.460 We need to have smaller chunks 00:47.460 --> 00:50.430 so we won't pass the token limitation. 00:50.430 --> 00:53.490 So, while this task may seem trivial, 00:53.490 --> 00:56.670 and we saw this in the course, it's often nuanced 00:56.670 --> 00:59.520 and overlooked because when splitting the text, 00:59.520 --> 01:02.520 we want to ensure that each chunk 01:02.520 --> 01:07.410 has some cohesive information that we can understand 01:07.410 --> 01:09.300 and the LLM can understand. 01:09.300 --> 01:11.550 We don't want to simply split a sentence 01:11.550 --> 01:12.930 in the middle of it. 01:12.930 --> 01:16.260 So, we want to have a chunk that is small enough 01:16.260 --> 01:18.120 that we can understand it, 01:18.120 --> 01:19.770 but not too big 01:19.770 --> 01:20.760 and not too large, 01:20.760 --> 01:23.700 so that it will have a lot of tokens. 01:23.700 --> 01:26.670 So, I often get messages of questions 01:26.670 --> 01:30.330 of how to split the data, how to set the chunk overlap, 01:30.330 --> 01:33.690 and what should be the chunk size, or what should we split. 01:33.690 --> 01:36.810 And to be honest, there is no correct answer for this, 01:36.810 --> 01:39.420 and every case needs to be examined 01:39.420 --> 01:41.370 and need to be handled differently. 01:41.370 --> 01:44.670 However, luckily for us, lang chain created a tool 01:44.670 --> 01:48.690 that we can use to visualize how we split our text. 01:48.690 --> 01:51.690 So, this is called the Text Splitter Playground. 01:51.690 --> 01:53.280 So, let me show how to find it. 01:53.280 --> 01:55.950 I'm simply going to Google lang chain 01:55.950 --> 01:57.633 text splitting playground, 01:59.280 --> 02:02.010 and I'm going to select the second result. 02:02.010 --> 02:04.050 This is the Git Hub repo. It's open source. 02:04.050 --> 02:05.700 You can check out this code. 02:05.700 --> 02:07.830 And this is basically a streaming publication. 02:07.830 --> 02:09.360 You can click on this link 02:09.360 --> 02:11.460 because it's hosted on lang chain. 02:11.460 --> 02:13.200 It's very intuitive to use. 02:13.200 --> 02:14.940 We can see at the first part, 02:14.940 --> 02:17.100 we can play around with the parameters. 02:17.100 --> 02:19.770 So, we can play around with the chunk size, 02:19.770 --> 02:21.420 with the chunk overlap, 02:21.420 --> 02:24.390 and how do we calculate this chunk size. 02:24.390 --> 02:27.000 And we can even select our text splitter. 02:27.000 --> 02:28.260 So, right now we're using 02:28.260 --> 02:31.200 the most common recursive character text splitter 02:31.200 --> 02:32.850 that we used in the course. 02:32.850 --> 02:35.940 And at the bottom, each time we change this parameter, 02:35.940 --> 02:37.980 we can see it reflected in the code, 02:37.980 --> 02:42.030 which we can simply copy paste into our workspace. 02:42.030 --> 02:43.440 Let's see this in action. 02:43.440 --> 02:45.720 And let me go to the lang chain blog 02:45.720 --> 02:49.980 and select a blog that we can copy data from. 02:49.980 --> 02:52.230 And let's see how we chunk it up. 02:52.230 --> 02:55.420 So, I'm simply going to copy all the text 02:56.760 --> 03:01.760 and I'm going to go back into the text splitter, pasting it. 03:01.980 --> 03:03.390 Let's click split text 03:03.390 --> 03:06.750 and let's take a look at the chunks. 03:06.750 --> 03:09.240 We can see visually all the chunks 03:09.240 --> 03:13.170 and we can make if they make sense or not. 03:13.170 --> 03:17.040 This way we can optimize our chunking strategy 03:17.040 --> 03:20.970 and to actually visualize and see how our chunks looks. 03:20.970 --> 03:23.040 And those are the chunks which are going 03:23.040 --> 03:26.910 to be embedded at the vector store, for example. 03:26.910 --> 03:28.320 This is also a good way 03:28.320 --> 03:30.210 to check out the chunk overlap. 03:30.210 --> 03:33.240 So, for example, we can see between two chunks, 03:33.240 --> 03:37.710 what is overlapped according to the overlapping size. 03:37.710 --> 03:39.330 And to wrap up this video, 03:39.330 --> 03:41.520 I really think this is a wonderful tool, 03:41.520 --> 03:43.860 which is helpful when we want 03:43.860 --> 03:46.440 to optimize our chunking strategy 03:46.440 --> 03:48.840 and when we want to visualize our chunks 03:48.840 --> 03:50.853 and what data they hold.