WEBVTT 1 00:00:00.000 --> 00:00:04.800 Hi, and welcome to another AI in C-Sharp video on the Microsoft Agent Framework. 2 00:00:05.599 --> 00:00:09.800 Today we're going to do a deep dive into chat reducers 3 00:00:09.800 --> 00:00:14.000 because we have made a previous video before, links in the description, 4 00:00:15.000 --> 00:00:17.600 about the two built-in chat reducers. 5 00:00:17.600 --> 00:00:20.000 But what if we want to do something different 6 00:00:20.000 --> 00:00:24.600 than just reduce the number of messages or do a simple summary? 7 00:00:25.799 --> 00:00:29.500 We are able to make our own custom reducers, 8 00:00:29.500 --> 00:00:31.500 and that's what we're going to cover today. 9 00:00:33.500 --> 00:00:39.099 So we have the previous sample here, chat history reducer, 10 00:00:39.700 --> 00:00:42.500 and the link to the video about that, 11 00:00:43.500 --> 00:00:47.900 where we went in and used the message counting reducer 12 00:00:47.900 --> 00:00:51.700 and the summary reducer, meaning four messages, 13 00:00:52.299 --> 00:00:54.099 get down to four messages again, 14 00:00:54.500 --> 00:00:57.500 and summary after four messages, summarized. 15 00:00:57.500 --> 00:01:03.500 But now we're going to go in and see not the built-in ones, 16 00:01:04.099 --> 00:01:06.099 but the built-in ones, as you can see here, 17 00:01:06.099 --> 00:01:09.300 is just an iChatReducer interface. 18 00:01:10.300 --> 00:01:15.099 So we can make our own, and then the first one I'm going to make here, 19 00:01:16.900 --> 00:01:20.099 or let's run it first so we can actually see 20 00:01:21.099 --> 00:01:24.500 how it works without any chat reducers. 21 00:01:25.099 --> 00:01:28.699 That might help a little with the understanding of this. 22 00:01:29.699 --> 00:01:31.300 So if I say hello to it, 23 00:01:33.300 --> 00:01:36.099 this sample will just write out the answer, 24 00:01:36.500 --> 00:01:38.500 how many number of messages there were, 25 00:01:38.500 --> 00:01:41.099 and who wrote them, user or assistant. 26 00:01:41.699 --> 00:01:47.500 And I also put in a tool so we can see tools in this. 27 00:01:47.699 --> 00:01:53.099 What is the weather like in Paris? 28 00:01:55.500 --> 00:01:58.099 And because we have a session on, 29 00:01:58.099 --> 00:02:01.099 we can of course see that both say hello, hello back, 30 00:02:01.500 --> 00:02:02.699 what is the weather like in Paris, 31 00:02:02.699 --> 00:02:05.699 then we get the tool call for get weather, 32 00:02:06.099 --> 00:02:10.100 and the tool result back, and then the final result. 33 00:02:10.500 --> 00:02:12.100 And if we keep doing this, 34 00:02:15.899 --> 00:02:20.300 we will see that this just builds up. 35 00:02:21.300 --> 00:02:25.699 And this is of course good for the AI, 36 00:02:25.699 --> 00:02:28.100 because now we could go in and say, 37 00:02:29.500 --> 00:02:33.899 what cities have we talked about? 38 00:02:42.899 --> 00:02:44.899 And it would come back and say, 39 00:02:44.899 --> 00:02:47.300 we have two cities, Paris and Copenhagen. 40 00:02:47.300 --> 00:02:50.699 And it can only do that because it has the entire history. 41 00:02:52.100 --> 00:02:55.500 But there could be reasons for getting rid of this history, 42 00:02:55.500 --> 00:02:59.899 mostly tokens and latency, 43 00:02:59.899 --> 00:03:02.899 because the more we send, the more there will be here. 44 00:03:03.699 --> 00:03:06.500 So the first one I want to introduce you to 45 00:03:06.500 --> 00:03:10.100 is one I made all back in Semantic Kernel, 46 00:03:10.100 --> 00:03:13.100 but I have now built here as well in Agent Framework. 47 00:03:13.500 --> 00:03:17.300 And that is one that always just removes all tool calls. 48 00:03:18.699 --> 00:03:21.699 So the way you do an iChat reducer 49 00:03:21.699 --> 00:03:25.899 is you just inherit from it or implement it, 50 00:03:25.899 --> 00:03:28.500 given that it's an interface. 51 00:03:28.699 --> 00:03:31.899 And the only thing you need to implement is a reduce async 52 00:03:31.899 --> 00:03:35.899 that takes in the previous messages and a cancellation token, 53 00:03:36.500 --> 00:03:40.500 and spits back what messages should be kept. 54 00:03:41.500 --> 00:03:45.500 So any messages we take out of this and put back in 55 00:03:45.500 --> 00:03:50.100 will be used and the rest will be gone. 56 00:03:51.500 --> 00:03:56.500 I'll put in a little utility just to see that it's been called. 57 00:03:57.100 --> 00:03:59.100 And then what are we doing here? 58 00:03:59.100 --> 00:04:02.699 We are making a list of the messages we want to keep 59 00:04:02.699 --> 00:04:04.699 from the previous messages. 60 00:04:05.300 --> 00:04:11.500 We are going in and say if that message is of role tool, 61 00:04:11.500 --> 00:04:14.100 then we get rid of it, meaning we just skip it, 62 00:04:14.100 --> 00:04:16.100 don't put it in to keep. 63 00:04:16.500 --> 00:04:19.100 Or if the message is from the assistant 64 00:04:19.100 --> 00:04:22.100 and it has a tool call request, 65 00:04:22.100 --> 00:04:25.100 meaning a function call content, 66 00:04:25.100 --> 00:04:27.100 then we also get rid of it. 67 00:04:27.700 --> 00:04:29.700 So let's see that as an action, 68 00:04:29.700 --> 00:04:31.899 in that we set a breakpoint here, 69 00:04:31.899 --> 00:04:35.899 and we go up and say that instead of no reducers, 70 00:04:35.899 --> 00:04:40.899 we want to use the always remove tool calls reducer. 71 00:04:43.299 --> 00:04:45.299 So let's run. 72 00:04:46.700 --> 00:04:49.700 Put this side by side so we can see. 73 00:04:52.100 --> 00:04:55.100 So if I just say hello, 74 00:04:57.500 --> 00:05:00.500 we can see it's called a reducer, 75 00:05:00.500 --> 00:05:03.500 and in the first case, because this is the first message, 76 00:05:04.100 --> 00:05:06.899 our previous message is actually zero. 77 00:05:06.899 --> 00:05:10.899 So technically it makes no sense to go to this 78 00:05:11.299 --> 00:05:15.299 because we just get to keep zero out of zero. 79 00:05:17.899 --> 00:05:21.899 And it will come back. Hello, how can I assist you today? 80 00:05:22.500 --> 00:05:27.500 So if I say, how are you? 81 00:05:27.899 --> 00:05:31.899 Again, we are called here. 82 00:05:32.700 --> 00:05:34.700 We get now two messages, 83 00:05:34.700 --> 00:05:37.100 so we don't get the third one, how are you, 84 00:05:37.100 --> 00:05:41.500 but we only get the two prior to what we're just about to call. 85 00:05:42.100 --> 00:05:46.100 And again, this happens before we get a message. 86 00:05:46.899 --> 00:05:50.899 So the two ones we have here is hello and hello back. 87 00:05:52.700 --> 00:05:55.700 So we call again the tool, and what to keep? 88 00:05:56.299 --> 00:05:59.299 Well, the first message is just hello, 89 00:05:59.700 --> 00:06:02.700 so it shouldn't be removed, 90 00:06:02.700 --> 00:06:05.700 and the second message is also just a response back. 91 00:06:07.100 --> 00:06:10.100 But it's your logic in C-sharp, 92 00:06:10.100 --> 00:06:12.700 so you can do pretty much anything here. 93 00:06:12.700 --> 00:06:14.700 You'll see a few other examples. 94 00:06:15.299 --> 00:06:18.299 So it came back, I'm just a programmer. 95 00:06:18.700 --> 00:06:21.700 So now we can, for example, say, 96 00:06:21.700 --> 00:06:27.700 what is the weather like in Paris? 97 00:06:30.299 --> 00:06:34.299 And for context here, I just gave the agent a tool, 98 00:06:36.700 --> 00:06:39.700 get weather, which just say it's sunny and 90 degrees, 99 00:06:39.700 --> 00:06:42.700 no matter what city it is, just for testing. 100 00:06:43.700 --> 00:06:47.700 So if we call this, again, we hit the reducer, 101 00:06:47.700 --> 00:06:50.700 we have four messages now, the hello and how are you, 102 00:06:52.299 --> 00:06:55.299 and we're still not going to remove anything, 103 00:06:55.700 --> 00:06:58.700 because the tool calls have not happened yet. 104 00:07:00.500 --> 00:07:04.500 So it always happens prior to the call, not after the call, 105 00:07:05.700 --> 00:07:08.700 because that could essentially remove some of the new stuff. 106 00:07:10.700 --> 00:07:14.700 So we're going to them, we see that we keep all four, 107 00:07:15.100 --> 00:07:18.100 because none of them are tool-related. 108 00:07:19.100 --> 00:07:24.100 And it comes back, and we can now see we have eight messages in here. 109 00:07:24.100 --> 00:07:27.100 So we saw the four before, and then we said, 110 00:07:27.100 --> 00:07:29.100 what is the weather like in Paris? 111 00:07:29.100 --> 00:07:32.100 We got a tool call for get weather, 112 00:07:32.100 --> 00:07:36.100 and we got a tool result back for what it was, 113 00:07:36.100 --> 00:07:41.100 and then the assistant came back and summarized these two things. 114 00:07:42.100 --> 00:07:48.100 So now we can try and say, what about Copenhagen? 115 00:07:51.100 --> 00:07:54.100 Meaning the weather in Copenhagen. 116 00:07:55.100 --> 00:07:58.100 So now we get our eight messages, 117 00:07:58.100 --> 00:08:02.100 but now we can see that at some points, 118 00:08:02.100 --> 00:08:05.100 let me just set some breakpoints here, 119 00:08:05.100 --> 00:08:09.100 we get a message that was the tool call, 120 00:08:10.100 --> 00:08:13.100 meaning it was this one here. 121 00:08:14.100 --> 00:08:18.100 So where it asks, I need to call the get weather tool, 122 00:08:18.100 --> 00:08:21.100 and we're now just skipping that, 123 00:08:21.100 --> 00:08:26.100 and we also get the tool back, meaning this one here, 124 00:08:27.100 --> 00:08:29.100 and we're also skipping that. 125 00:08:29.100 --> 00:08:31.100 So if we set a breakpoint up here, 126 00:08:31.100 --> 00:08:34.099 we now have six messages we want to return, 127 00:08:35.099 --> 00:08:37.099 and that is the hello, 128 00:08:37.099 --> 00:08:41.099 so the back and forward that we have really seen, 129 00:08:41.099 --> 00:08:43.099 but the tool call is now gone. 130 00:08:44.099 --> 00:08:48.099 And we need to remove both the assistant request and the tool, 131 00:08:48.099 --> 00:08:52.099 else the AI will actually break down and say, 132 00:08:52.099 --> 00:08:57.099 there isn't a matching pair of request and result of this tool. 133 00:08:59.099 --> 00:09:02.099 But if we do this, and you can just copy this 134 00:09:02.099 --> 00:09:04.099 if you want to have this, 135 00:09:05.099 --> 00:09:09.099 we now see that that is what is being sent prior, 136 00:09:10.099 --> 00:09:16.099 and we get rid of the tool calls up here for the ask of Paris, 137 00:09:16.099 --> 00:09:20.099 but we of course see the new tool call down here 138 00:09:20.099 --> 00:09:22.099 for the new message. 139 00:09:22.099 --> 00:09:25.099 But when we send the next one, it will remove those two as well. 140 00:09:25.099 --> 00:09:28.099 So this one will always remove all tools, 141 00:09:28.099 --> 00:09:35.099 but the tools that were used in the last sent message. Okay. 142 00:09:36.099 --> 00:09:38.099 Let's try something else. 143 00:09:40.099 --> 00:09:43.099 Let's try to just make a thing 144 00:09:43.099 --> 00:09:47.099 that gets rid of the word sunny in messages, 145 00:09:49.099 --> 00:09:53.099 because they don't need to be as big as all this up here. 146 00:09:53.099 --> 00:09:57.099 Here, it can simply see to remove all previous messages 147 00:09:57.099 --> 00:10:03.099 that contains the word sunny in the invariant case. 148 00:10:05.099 --> 00:10:10.099 So if we do that, let me get rid of this one, so we can see. 149 00:10:10.099 --> 00:10:13.099 Essentially, we're just taking the previous messages 150 00:10:13.099 --> 00:10:15.099 and filtering them a bit. 151 00:10:17.099 --> 00:10:21.099 That could be filtering for some words that shouldn't be there 152 00:10:21.099 --> 00:10:26.099 or out-of-date information or whatever it is. 153 00:10:26.099 --> 00:10:31.099 But if we take these messages with word reducer 154 00:10:31.099 --> 00:10:36.099 and set it to be the chat reducer, and restart. 155 00:10:39.099 --> 00:10:44.099 So now we can again say, what is the weather like in Paris? 156 00:10:45.099 --> 00:10:49.099 And we will get our four messages, 157 00:10:49.099 --> 00:10:53.099 the user message, the tool call request, 158 00:10:53.099 --> 00:10:56.099 the tool back, and the assistant. 159 00:10:58.099 --> 00:11:02.099 And we can see that it mentioned the word sunny. 160 00:11:03.099 --> 00:11:07.099 So if we say, how about Copenhagen? 161 00:11:09.099 --> 00:11:12.099 It might surprise you what you will see back. 162 00:11:12.099 --> 00:11:14.099 We will see that we have seven messages. 163 00:11:14.099 --> 00:11:16.099 We have the user. 164 00:11:16.099 --> 00:11:18.099 We have the assistant. 165 00:11:20.099 --> 00:11:23.099 We have the sunny and 19 degrees. 166 00:11:23.099 --> 00:11:27.099 But we said we wanted to get rid of those. 167 00:11:27.099 --> 00:11:33.099 Well, down here, I only... Let me find it. 168 00:11:33.099 --> 00:11:36.099 I only set the one that had text in them. 169 00:11:36.099 --> 00:11:39.099 And this is not text, this is content. 170 00:11:39.099 --> 00:11:44.099 So if you want to get rid of everything that has anything to do with tools, 171 00:11:44.099 --> 00:11:49.099 you first need to check each tool if they mention the word sunny. 172 00:11:49.099 --> 00:11:55.099 So the only thing that is gone here is actually the first reply back. 173 00:11:56.099 --> 00:12:00.099 The weather in Paris is sunny and 19 degrees. That is gone. 174 00:12:00.099 --> 00:12:03.099 And now we only have the Copenhagen weather. 175 00:12:03.099 --> 00:12:07.099 So it's a bit contrived, but just to show you that 176 00:12:07.099 --> 00:12:12.099 whenever you do this, it is your job to get rid of whatever it is. 177 00:12:12.099 --> 00:12:15.099 And in this case, text is only the one 178 00:12:15.099 --> 00:12:18.099 that actually gets rid of the final assistant message. 179 00:12:20.099 --> 00:12:22.099 Let's try another one. 180 00:12:23.099 --> 00:12:30.099 We can also go in and use agents within the chat reducer. 181 00:12:30.099 --> 00:12:35.099 So in this case, I'm making a pirate summary reducer agent. 182 00:12:36.099 --> 00:12:39.099 So it's a little like the summary chat reducer up here, 183 00:12:39.099 --> 00:12:43.099 but we get more control. We can use tools and so on. 184 00:12:43.099 --> 00:12:45.099 We can technically do that up here, 185 00:12:45.099 --> 00:12:47.099 but it's not with the agent framework, 186 00:12:47.099 --> 00:12:51.099 it's with the underlying Microsoft Extensions AI. 187 00:12:52.099 --> 00:12:56.099 So in this case, I'm just making a pirate summary reducer 188 00:12:56.099 --> 00:12:59.099 and telling it, giving the input messages, 189 00:12:59.099 --> 00:13:02.099 make a summary of them in the voice of a pirate. 190 00:13:03.099 --> 00:13:07.099 And I'm saying it only needs to happen after four messages. 191 00:13:09.099 --> 00:13:15.099 So let's take this one, and then we'll see the code in one second. 192 00:13:18.099 --> 00:13:23.099 So we'll just go down here and set up. 193 00:13:27.099 --> 00:13:29.099 There we go. 194 00:13:30.099 --> 00:13:33.099 So we will try again to say, 195 00:13:33.099 --> 00:13:36.099 what is the weather like in Paris? 196 00:13:36.099 --> 00:13:40.099 We see that our tool is being called, 197 00:13:40.099 --> 00:13:44.099 but we will just say no summary yet 198 00:13:44.099 --> 00:13:47.099 because we only have zero messages 199 00:13:47.099 --> 00:13:52.099 and not the minimum four before we actually do a summarization. 200 00:13:53.099 --> 00:13:57.099 We can try again and say, what about Copenhagen? 201 00:13:58.099 --> 00:14:01.099 We only have four messages, 202 00:14:01.099 --> 00:14:05.099 and in this case we have set four or lower, 203 00:14:05.099 --> 00:14:08.099 so still no summarization. 204 00:14:11.099 --> 00:14:14.099 So we now have eight messages 205 00:14:14.099 --> 00:14:19.099 with the message from Paris and a message from Copenhagen. 206 00:14:19.099 --> 00:14:22.099 Let's say Berlin. 207 00:14:23.099 --> 00:14:26.099 We now have eight messages, 208 00:14:26.099 --> 00:14:29.099 so we have one to do summary. 209 00:14:29.099 --> 00:14:32.099 And we put in an agent, so we can just use that agent 210 00:14:32.099 --> 00:14:34.099 for any kind of thing we want. 211 00:14:34.099 --> 00:14:38.099 In this case, we gave it instructions to agent 212 00:14:38.099 --> 00:14:41.099 that given the input amount, 213 00:14:41.099 --> 00:14:44.099 summarizing the voice of a pirate. 214 00:14:44.099 --> 00:14:49.099 That given the input amount, summarizing the voice of a pirate. 215 00:14:49.099 --> 00:14:54.099 This is just an agent framework call like any other. 216 00:14:56.099 --> 00:14:59.099 So it will come back and, in this case, 217 00:14:59.099 --> 00:15:03.099 summarize those eight messages. 218 00:15:03.099 --> 00:15:07.099 Now we're only returning one message back. 219 00:15:08.099 --> 00:15:10.099 So we have essentially taken the eight messages, 220 00:15:10.099 --> 00:15:13.099 turned it into one message. 221 00:15:15.099 --> 00:15:18.099 So we will see that we have five messages. 222 00:15:18.099 --> 00:15:21.099 So we have the summarized message, 223 00:15:21.099 --> 00:15:24.099 we have the Berlin message, 224 00:15:24.099 --> 00:15:29.099 we have the AI requesting the tool, 225 00:15:30.099 --> 00:15:32.099 and the tool response, 226 00:15:32.099 --> 00:15:36.099 and back what actually happens. 227 00:15:36.099 --> 00:15:39.099 So that makes all sense. 228 00:15:39.099 --> 00:15:43.099 We normally get four messages every time we ask. 229 00:15:43.099 --> 00:15:45.099 One input, one output, 230 00:15:45.099 --> 00:15:50.099 and the two tool calling request and response. 231 00:15:50.099 --> 00:15:53.099 The fifth one is now our summary. 232 00:15:53.099 --> 00:15:55.099 And given that we have a summary now, 233 00:15:55.099 --> 00:15:58.099 we can go in and say, 234 00:15:58.099 --> 00:16:11.099 what cities have we talked about so far? 235 00:16:13.099 --> 00:16:15.099 Again, we get five messages, 236 00:16:15.099 --> 00:16:18.099 which is more than what we want, 237 00:16:18.099 --> 00:16:23.099 so we actually do a summarization of this prior. 238 00:16:24.099 --> 00:16:29.099 So it will talk about all the cities and so on. 239 00:16:29.099 --> 00:16:32.099 We will send that. 240 00:16:32.099 --> 00:16:33.099 And it will come and say, 241 00:16:33.099 --> 00:16:38.099 the cities mentioned so far are Paris, Copenhagen, and Berlin. 242 00:16:40.099 --> 00:16:42.099 And now we're down to three messages 243 00:16:42.099 --> 00:16:46.099 because we only got the summary, 244 00:16:46.099 --> 00:16:48.099 the question, and the answer, 245 00:16:48.099 --> 00:16:57.099 because it didn't need to do any calls to the tools. 246 00:16:57.099 --> 00:17:00.099 Had we asked, the city mentioned so far, 247 00:17:00.099 --> 00:17:02.099 what are their weather? 248 00:17:02.099 --> 00:17:05.099 We can try and do it. 249 00:17:15.099 --> 00:17:18.099 Now we again do a summary. 250 00:17:19.099 --> 00:17:23.099 Now we're down to three, so we don't do a summary, of course, 251 00:17:23.099 --> 00:17:27.099 because we are below the four. 252 00:17:29.099 --> 00:17:32.099 And now we see a bunch of extra tool calls 253 00:17:32.099 --> 00:17:35.099 because it grabbed three tool calls, 254 00:17:35.099 --> 00:17:38.099 one for Paris, one for Copenhagen, one for Berlin. 255 00:17:38.099 --> 00:17:41.099 And the prior got the same message back for all of them 256 00:17:41.099 --> 00:17:44.099 because it's just hard-coded. 257 00:17:44.099 --> 00:17:49.099 But that makes sense that we can get multiple backs, 258 00:17:49.099 --> 00:17:54.099 and now we get the summary again, and so on and so forth. 259 00:17:54.099 --> 00:17:56.099 Of course, summary will cost extra 260 00:17:56.099 --> 00:18:00.099 because we do both in cost and time if we do an agent, 261 00:18:00.099 --> 00:18:03.099 but you could do pretty much anything here. 262 00:18:03.099 --> 00:18:06.099 And I've tried to do one more of these 263 00:18:06.099 --> 00:18:10.099 where I do a city reducer agent. 264 00:18:10.099 --> 00:18:13.099 So given the input number messages, 265 00:18:13.099 --> 00:18:18.099 return the numbers of the messages that contains a city name. 266 00:18:19.099 --> 00:18:24.099 So this one, let's put it down here and have a look at it. 267 00:18:25.099 --> 00:18:29.099 This is a bit more contrived, but just for you to see. 268 00:18:29.099 --> 00:18:32.099 So if we have no messages, we don't need to do anything, 269 00:18:32.099 --> 00:18:35.099 but if we do have any messages, 270 00:18:35.099 --> 00:18:40.099 we can go in and first use one of the other reducers, 271 00:18:40.099 --> 00:18:44.099 meaning I just get rid of all the tool calls on the fly 272 00:18:44.099 --> 00:18:46.099 by doing reduce down here. 273 00:18:46.099 --> 00:18:48.099 This is the best way of chaining it 274 00:18:48.099 --> 00:18:51.099 because you can't chain them up in Agent Framework yourself, 275 00:18:51.099 --> 00:18:55.099 but you can easily just new up one of the other reducers 276 00:18:55.099 --> 00:18:59.099 and quickly call reduce before you work with the data. 277 00:19:00.099 --> 00:19:06.099 And then I'm building up a message of all the messages in string 278 00:19:06.099 --> 00:19:08.099 and send them into the system, 279 00:19:08.099 --> 00:19:14.099 so it can give me back what numbered lines are to be containing a city 280 00:19:14.099 --> 00:19:16.099 and then get rid of it. 281 00:19:16.099 --> 00:19:19.099 So the best way to understand that one 282 00:19:19.099 --> 00:19:23.099 is probably to set a breakpoint here and restart. 283 00:19:24.099 --> 00:19:29.099 So if we say, what is the weather like in Paris, 284 00:19:30.099 --> 00:19:32.099 we of course are called, that's zero, 285 00:19:32.099 --> 00:19:35.099 so we don't need to do anything, 286 00:19:35.099 --> 00:19:38.099 and push it back. 287 00:19:40.099 --> 00:19:42.099 How about Copenhagen? 288 00:19:42.099 --> 00:19:48.099 So now we see that we have a user message, of course, that is Paris. 289 00:19:48.099 --> 00:19:52.099 We have a couple of tool calls, 290 00:19:52.099 --> 00:19:55.099 and we have the word Paris here again. 291 00:19:55.099 --> 00:20:00.099 So if we do this, try and quickly stop the video 292 00:20:00.099 --> 00:20:05.099 and think what would you expect comes back. 293 00:20:08.099 --> 00:20:11.099 Okay, let's try and do it. 294 00:20:11.099 --> 00:20:15.099 So we have four previous messages, 295 00:20:15.099 --> 00:20:20.099 so we say we will always get rid of tool calls, 296 00:20:21.099 --> 00:20:24.099 meaning we call that one. 297 00:20:24.099 --> 00:20:28.099 Let me quickly get out of it, so we go down here. 298 00:20:29.099 --> 00:20:34.099 So now we have just taken the four messages 299 00:20:34.099 --> 00:20:37.099 and reduced them to two. 300 00:20:38.099 --> 00:20:41.099 And that is the original one and the real answer. 301 00:20:41.099 --> 00:20:43.099 So the tool calls are gone. 302 00:20:43.099 --> 00:20:48.099 And now we build a small string builder here. 303 00:20:48.099 --> 00:20:54.099 So what we are saying is, zero is this line, one is this line. 304 00:20:55.099 --> 00:20:57.099 Then I'm doing structured output to tell, 305 00:20:57.099 --> 00:21:02.099 give me all the lines that actually have a city named. 306 00:21:03.099 --> 00:21:09.099 In this case, it's both the line index zero and index one. 307 00:21:10.099 --> 00:21:15.099 So given that both of these contain the word Paris, 308 00:21:16.099 --> 00:21:19.099 we actually remove all messages. 309 00:21:20.099 --> 00:21:24.099 And that results in, if we have a look, 310 00:21:25.099 --> 00:21:28.099 the weather in Copenhagen is currently sunny 311 00:21:28.099 --> 00:21:31.099 with the temperature of 90 degrees. 312 00:21:31.099 --> 00:21:37.099 So the reason why it knows this is because it, 313 00:21:37.099 --> 00:21:39.099 what about Copenhagen? 314 00:21:39.099 --> 00:21:43.099 And it had a getWeather, but it had no context of weather. 315 00:21:43.099 --> 00:21:46.099 So this is just the AI being clever enough that 316 00:21:46.099 --> 00:21:50.099 what about Copenhagen is getWeather. 317 00:21:51.099 --> 00:21:54.099 But had we said hello prior to this, 318 00:21:54.099 --> 00:21:57.099 and then ask for this, let's try to do it. 319 00:21:58.099 --> 00:22:03.099 Let me get rid of all these breakpoints so we can just see it. 320 00:22:04.099 --> 00:22:07.099 So let's say first hi to it. 321 00:22:11.099 --> 00:22:13.099 That comes back. 322 00:22:13.099 --> 00:22:16.099 What's the weather like in Paris? 323 00:22:17.099 --> 00:22:19.099 That comes back, which is okay. 324 00:22:20.099 --> 00:22:23.099 And what about Copenhagen? 325 00:22:24.099 --> 00:22:29.099 Now we can see that the hi is there, the how back, 326 00:22:29.099 --> 00:22:32.099 and then we get to what about Copenhagen. 327 00:22:33.099 --> 00:22:37.099 And now you can see it's a bit unsure. 328 00:22:37.099 --> 00:22:41.099 Are you talking about the weather or is it something else? 329 00:22:41.099 --> 00:22:44.099 Because the context was lost because all the messages 330 00:22:44.099 --> 00:22:49.099 that had a city name involved was removed. 331 00:22:49.099 --> 00:22:56.099 Again, a bit contrived, but again, the primary message here is 332 00:22:56.099 --> 00:23:00.099 you have full control to do anything you want in here. 333 00:23:00.099 --> 00:23:03.099 You just get, here are some messages, 334 00:23:03.099 --> 00:23:05.099 which of them do I want to keep? 335 00:23:05.099 --> 00:23:09.099 You can use AI, you can use regex, you can use common sense, 336 00:23:09.099 --> 00:23:15.099 you can use counting, anything you can do to your imagination, 337 00:23:15.099 --> 00:23:18.099 you can do here in order to get rid of that. 338 00:23:19.099 --> 00:23:22.099 So that's everything. See you in the next one.