WEBVTT 00:00.120 --> 00:05.720 Messages are the fundamental units of communicating with llms. 00:05.760 --> 00:14.120 When working with chat models in Linkchain, they carry the input we want to send to the LM and the 00:14.120 --> 00:16.120 output we receive back. 00:16.560 --> 00:19.680 Every message has two essential parts. 00:19.680 --> 00:23.160 It has a role and it has content. 00:23.680 --> 00:26.680 The role tells the model who sent the message. 00:27.040 --> 00:33.200 The content holds the actual information, which is typically text, but can also include things like 00:33.200 --> 00:37.680 images, videos, or other data for multimodal models. 00:38.000 --> 00:45.400 Linkchain standardizes these messages, allowing us to work with different chat model provider using 00:45.400 --> 00:47.120 a consistent format. 00:47.240 --> 00:50.880 So this means it doesn't matter which LM we're using. 00:51.080 --> 00:55.200 We still use the same interface, which is very, very convenient. 00:55.600 --> 00:59.320 So let's take a look on the different roles messages can have. 00:59.680 --> 01:06.570 We have the system role which is used to set up the AI's behavior or to provide initial context. 01:06.610 --> 01:13.330 For instance, you might use a system message to tell the model to act as a helpful assistant or to 01:13.370 --> 01:15.970 summarize complex documents concisely. 01:16.490 --> 01:19.410 Not all models handle system message, by the way. 01:19.690 --> 01:21.970 Some use a dedicated parameter. 01:22.010 --> 01:24.770 Other integrated into the message history. 01:24.970 --> 01:27.170 And few don't support it at all. 01:27.210 --> 01:34.450 However, Linkchain tries to handle these variations for us and it's going to create one single interface. 01:35.410 --> 01:43.610 The user role represents input coming from a person or system interacting with the model in Linkchain. 01:43.650 --> 01:46.330 We'll use the human message class for this. 01:46.850 --> 01:55.050 When we send text to a chat model, like asking a question or giving an instruction, we wrap that content 01:55.050 --> 01:57.210 in a human message class. 01:57.530 --> 02:00.170 Linkchain even offers us a shortcut. 02:00.210 --> 02:06.980 If we want to invoke a model with just a string, it's automatically being treated as a human message. 02:07.860 --> 02:10.660 The assistant role is the model's response. 02:11.100 --> 02:16.220 This output is representative by the AI message class in LinkedIn. 02:16.740 --> 02:25.340 When the model answers our question or completes a task, its response comes back as an AI message object. 02:25.660 --> 02:33.860 These messages contain the main content generated by the large language model, along with other information 02:33.860 --> 02:41.300 like metadata or tool calls, token usage information, and unique identifiers to help us in debugging. 02:41.900 --> 02:46.380 Tool role will elaborate this when we talk about agents and function calling. 02:46.580 --> 02:52.380 But overall, it's a way for us to tell the AI what was the result of the tool execution. 02:52.460 --> 02:59.060 For example, if we invoked a tool that gets us the weather in somewhere, then we're going to wrap 02:59.060 --> 03:02.540 that answer in a tool message object. 03:02.540 --> 03:06.760 And that's how we're going to tell the model that that's the result of the tool call. 03:06.760 --> 03:08.840 So it can give us the final answer. 03:08.840 --> 03:11.360 But don't worry if you don't fully understand this part. 03:11.400 --> 03:15.280 We're going to elaborate a lot on this topic in the course. 03:16.200 --> 03:24.680 The order of these messages is important for a coherent conversation, and a typical flow involves alternating 03:24.680 --> 03:27.880 between user input and assistant responses. 03:28.200 --> 03:34.640 So it's going to be human message, then AI message, then human message, then AI message, and so 03:34.640 --> 03:34.920 on. 03:35.400 --> 03:42.920 If tools are involved, we probably see them in like and if tools are involved, we'll see this in human 03:42.920 --> 03:43.560 message. 03:43.600 --> 03:51.000 AI message with a tool call, then a tool message with the tool result, then an AI message using the 03:51.000 --> 03:53.160 tool result, and so on and so on. 03:53.200 --> 04:01.000 And just to conclude, link Messages object provide us with a consistent way to build these conversations, 04:01.040 --> 04:06.800 abstracting away the specific message formats required by different providers.