WEBVTT 00:00.000 --> 00:02.490 -: In this video we're gonna have a look at a variety 00:02.490 --> 00:04.380 of different approaches to summarizing, 00:04.380 --> 00:07.170 specifically looking at summarizing chains. 00:07.170 --> 00:08.940 Now we'll do our regular form of import, 00:08.940 --> 00:11.580 so Chat OpenAI, LLM chains. 00:11.580 --> 00:12.690 There's some new ones here. 00:12.690 --> 00:15.150 So for example, we can use a document loader 00:15.150 --> 00:18.630 to specifically load a URL or a list of URLs. 00:18.630 --> 00:20.460 You're also gonna need to import something called 00:20.460 --> 00:22.200 a load summarize chain 00:22.200 --> 00:25.770 and so when we run this, you'll see what happens is 00:25.770 --> 00:27.870 basically we load in some documents 00:27.870 --> 00:30.750 and these get assigned to LangChain documents. 00:30.750 --> 00:33.480 We then load up a chat open AI model, 00:33.480 --> 00:37.920 and then we use a load summarize chain called a stuff chain, 00:37.920 --> 00:40.830 which basically means that we stuff all of the documents 00:40.830 --> 00:43.080 into a single LLM call. 00:43.080 --> 00:44.130 That's a really important point. 00:44.130 --> 00:48.120 So again, stuffing is putting all of your LangChain 00:48.120 --> 00:50.910 documents into a single call, 00:50.910 --> 00:53.970 which does mean that if you have more documents than you can 00:53.970 --> 00:57.030 fit into a model, it will return an error. 00:57.030 --> 00:58.350 And you can see after we've done that, 00:58.350 --> 01:00.180 we've got a nice little summary explaining 01:00.180 --> 01:03.101 that this article is about building autonomous agents, 01:03.101 --> 01:05.640 which are powered by these large language models. 01:05.640 --> 01:08.400 We have then done a different approach, 01:08.400 --> 01:11.700 which is using something called the map reduce pattern. 01:11.700 --> 01:14.398 So you import a couple of different variables 01:14.398 --> 01:16.680 and things like the reduced documents, 01:16.680 --> 01:20.070 a map reduced documents chain and a stuff documents chain. 01:20.070 --> 01:21.870 You first have to set up your map, 01:21.870 --> 01:25.913 which is basically mapping over a set of documents 01:25.913 --> 01:29.730 and basically on that, identifying the main themes. 01:29.730 --> 01:34.320 And then you create a map chain whilst the reduce takes all 01:34.320 --> 01:37.530 of the summaries that are created inside of your map. 01:37.530 --> 01:40.260 And what it will do is it will then reduce those summaries 01:40.260 --> 01:42.030 down into a single summary. 01:42.030 --> 01:44.430 So map reduce runs really, really quickly. 01:44.430 --> 01:46.860 It's one of the faster ways of doing summaries. 01:46.860 --> 01:48.720 The problem with the map reduce chain 01:48.720 --> 01:51.000 is the pages don't remember each other. 01:51.000 --> 01:55.440 So, it can cause an issue where if page three kind of needs 01:55.440 --> 01:56.580 to know about page five, 01:56.580 --> 01:58.560 then the summaries are all done independently. 01:58.560 --> 02:01.740 So you can get some different forms of summarization 02:01.740 --> 02:02.820 which will handle for that. 02:02.820 --> 02:05.310 Moving back up, we use this LLM chain 02:05.310 --> 02:08.490 and we run something called a stuff documents chain, 02:08.490 --> 02:10.400 which takes in the reduced chain. 02:10.400 --> 02:13.770 And it's basically what it does is it combines 02:13.770 --> 02:16.770 these documents into a single string, right? 02:16.770 --> 02:18.420 So it takes a list of documents 02:18.420 --> 02:20.343 and combines 'em into a single string. 02:21.810 --> 02:24.120 And then what we do is we build something called 02:24.120 --> 02:26.880 a reduced documents chain, which combines 02:26.880 --> 02:29.520 and iteratively reduces the mapped documents. 02:29.520 --> 02:33.180 So the reduced chain here is this one, 02:33.180 --> 02:35.280 which is your stuffed documents chain. 02:35.280 --> 02:38.400 And then we have a collapsed documents chain, which is set 02:38.400 --> 02:40.080 to equal to the combined documents chain. 02:40.080 --> 02:42.093 So we're just using the same chain here. 02:42.093 --> 02:44.333 And this reduced documents chain, 02:44.333 --> 02:46.630 what this basically does is 02:47.520 --> 02:51.420 it's gonna reduce your documents down using map reduce. 02:51.420 --> 02:52.253 All right? 02:52.253 --> 02:55.290 And the map prompt comes from up here, right? 02:55.290 --> 02:57.270 So this is where the map chain is used. 02:57.270 --> 02:59.760 And then we have this map reduce documents chain, 02:59.760 --> 03:00.593 which is the final bit. 03:00.593 --> 03:03.630 So this is the signing, the reduce chain, right, 03:03.630 --> 03:06.690 which uses this reduce template. 03:06.690 --> 03:09.900 And then after that then we've got the reduce chain set up. 03:09.900 --> 03:11.490 We then make the map reduce chain. 03:11.490 --> 03:14.430 So the LLM chain that we're gonna use for our map chain 03:14.430 --> 03:15.840 is this one here. 03:15.840 --> 03:20.100 The reduce chain is this one which reduces the documents. 03:20.100 --> 03:23.100 And this is the variable name in the LLM chain, 03:23.100 --> 03:26.820 which we call here in the map, which is called docs. 03:26.820 --> 03:29.400 And we do also return immediate steps. 03:29.400 --> 03:31.380 And so when we run that, 03:31.380 --> 03:35.670 it gives us a summary using map reduce, all right? 03:35.670 --> 03:37.140 So you can see here we've got a couple 03:37.140 --> 03:39.570 of different summaries that have been therefore reduced. 03:39.570 --> 03:40.403 So there you go. 03:40.403 --> 03:41.910 So that's map reduce. 03:41.910 --> 03:44.610 The final summary type that we have is something called 03:44.610 --> 03:48.120 a refine template, which is basically like where you say, 03:48.120 --> 03:49.705 right, a concise summary of the following 03:49.705 --> 03:52.440 and you ask it to have a summary. 03:52.440 --> 03:55.380 And then we have a something called a refined template. 03:55.380 --> 03:56.490 And this refined template 03:56.490 --> 03:59.220 is to basically take the existing summary 03:59.220 --> 04:02.130 that get generated from these, provide it new texts 04:02.130 --> 04:04.170 and say, given the new context, 04:04.170 --> 04:05.400 could you do something to that? 04:05.400 --> 04:07.680 So could you refine the original summary, 04:07.680 --> 04:10.350 so we could do this, refine the original summary, 04:10.350 --> 04:13.050 if the content isn't useful, return the original summary. 04:13.050 --> 04:16.080 And then when we do our load summarization chain, 04:16.080 --> 04:18.810 what we're doing is we're passing in the question prompt, 04:18.810 --> 04:21.300 which comes from this prompt template here. 04:21.300 --> 04:24.570 And then we also have our refined prompt, 04:24.570 --> 04:26.100 which comes from here. 04:26.100 --> 04:30.060 And then we put our input key, which is our input documents 04:30.060 --> 04:33.930 and our output key, which is our output text. 04:33.930 --> 04:36.450 And then we'll end up with a result. 04:36.450 --> 04:39.840 And what's good about the refined way is if you want 04:39.840 --> 04:43.110 to essentially read the document, one page, 04:43.110 --> 04:45.900 then the second page, and then page three and page four. 04:45.900 --> 04:49.410 That's what the refined summarization technique is doing. 04:49.410 --> 04:52.950 It's essentially saying, let's start with page one. 04:52.950 --> 04:54.030 We'll make a summary. 04:54.030 --> 04:56.670 We'll go onto page two, refine, 04:56.670 --> 04:59.460 and then we'll go onto page three and then we'll refine. 04:59.460 --> 05:02.910 So the danger of using refine is that it's gonna take you 05:02.910 --> 05:05.545 a lot longer than something like MapReduce, 05:05.545 --> 05:09.060 because MapReduce runs each individual summarization 05:09.060 --> 05:13.718 independently whilst refine has to be a sequential series 05:13.718 --> 05:15.540 of steps, right? 05:15.540 --> 05:18.669 And those steps will lead to a greater summary. 05:18.669 --> 05:21.060 But the benefit of refine is that obviously 05:21.060 --> 05:23.190 it can improve the summary over time. 05:23.190 --> 05:25.650 So here's three different summarization techniques 05:25.650 --> 05:28.953 that you can use inside LangChain to get better results.