WEBVTT 00:00.620 --> 00:06.110 So this video is going to be all about memory and specifically all about memory in lecture one. 00:06.560 --> 00:10.880 As far as user interactions when it comes to LMS, LMS are stateless. 00:10.880 --> 00:16.760 So this means that the LMS don't save any information in the conversation that occurred in the conversation 00:16.760 --> 00:17.330 earlier. 00:17.330 --> 00:18.560 So it's stateless. 00:18.560 --> 00:20.090 Let's take a look at an example. 00:20.390 --> 00:26.000 Now if I'm going to ask my large language model that is based on the long chain documentation, who 00:26.000 --> 00:31.670 created long chain, then I would get a correct response that it was created by Harrison Chase. 00:31.670 --> 00:39.080 But if I'm going to follow up and ask, do you know any YouTube videos related to him, then the answer 00:39.080 --> 00:43.580 that I will get is, I'm sorry, I don't know who you're referring to. 00:43.610 --> 00:47.570 Could you please provide me with more context or clarify who is him? 00:47.930 --> 00:55.400 So the concept that we're discussing right now in the formal terminology is called a coreference resolution. 00:55.400 --> 01:03.410 Basically, it's the task of identifying all expressions, words or phrases in a text that refer to 01:03.440 --> 01:05.290 the same entity or concept. 01:05.290 --> 01:10.540 So in other words, it's the process of identifying all the instances where different words or phrases 01:10.540 --> 01:12.370 in a text refer to the same thing. 01:12.400 --> 01:16.000 And in our example, him referred to Harris and Chase. 01:16.030 --> 01:21.910 Now the LM was not able to do it because it didn't have a state, so it wasn't able to make a co-reference 01:21.910 --> 01:22.570 resolution. 01:22.570 --> 01:29.680 However, if the LM in the prompt will get the state and the chat history, then it will be easily able 01:29.680 --> 01:31.930 to make the co-reference resolution. 01:31.930 --> 01:37.240 And this is the underlying base to all the solutions that blockchain is currently supporting. 01:37.240 --> 01:43.720 For memory, we simply find sophisticated ways to pass into the prompt some data, some information 01:43.720 --> 01:49.270 that will help the LM to make some co-reference resolution, like in the example we see right now, 01:49.300 --> 01:52.930 the prompt that we're giving it is given the past conversation. 01:52.930 --> 01:54.550 Answer my question now. 01:54.550 --> 01:59.890 In the past conversation, we're simply going to answer an interaction between us and the bot. 01:59.890 --> 02:03.040 Something about I like drinking cold brew coffee. 02:03.070 --> 02:04.240 Where can I get it? 02:04.240 --> 02:07.860 And the last question is where else can I find that? 02:07.860 --> 02:11.880 And by that I am referencing the cold brew coffee. 02:11.940 --> 02:19.230 And also the LM will know that I don't want to drink at Starbucks, nor at Coffee Bean, and the LM 02:19.230 --> 02:22.890 can easily process that and can make some co-reference resolution. 02:22.950 --> 02:31.020 Now, you may notice that if we have a one hour conversation, then we are in a problem because we have 02:31.020 --> 02:32.820 too much data to put into the prompt. 02:32.820 --> 02:37.500 We know that there is a token limit, and if we have a very, very long conversation, then it will 02:37.500 --> 02:39.450 definitely exceed this limit. 02:40.620 --> 02:41.010 Okay. 02:41.010 --> 02:43.620 So now let's see how long chain resolves it. 02:43.620 --> 02:47.220 And let's take a look at all the long chain offerings regarding memory. 02:47.880 --> 02:50.460 Now this video is a theoretical video. 02:50.460 --> 02:54.570 So I'm not planning to show you any live examples of how to use them. 02:54.570 --> 02:58.740 They're pretty straightforward and we're going to see examples of them in the rest of this course. 02:58.740 --> 03:03.120 So don't worry about seeing implementations of those classes right now. 03:03.120 --> 03:07.980 What I want to demonstrate to you is the strategies that are used in order to resolve this token issue 03:08.010 --> 03:09.360 when it comes to memory.