1 00:00:00,040 --> 00:00:00,400 Okay. 2 00:00:00,440 --> 00:00:02,840 Next up let's work on the prompts. 3 00:00:02,960 --> 00:00:05,040 So we're going to improve the system message. 4 00:00:05,200 --> 00:00:06,920 And let me read you what we've got. 5 00:00:06,960 --> 00:00:10,320 And note again I'm using this trick of just using a global variable. 6 00:00:10,320 --> 00:00:12,960 So it's automatically going to apply everywhere earlier. 7 00:00:13,200 --> 00:00:17,280 So system message you're a helpful assistant in a clothes store. 8 00:00:17,440 --> 00:00:21,480 You should try to gently encourage the customer to try items that are on sale. 9 00:00:21,800 --> 00:00:26,640 Hats are 60% off and most other items are 60% or 50% off. 10 00:00:26,960 --> 00:00:32,240 For example, if the customer says, I'm looking to buy a hat, you could reply with something like 11 00:00:32,240 --> 00:00:32,720 wonderful! 12 00:00:32,720 --> 00:00:36,800 We have lots of hats, including several that are part of our sales event. 13 00:00:37,080 --> 00:00:42,280 Encourage the customer to buy hats if they're unsure what to get. 14 00:00:42,840 --> 00:00:45,720 Okay, so there's a lot of things going on in that system prompt. 15 00:00:45,760 --> 00:00:47,440 We're giving it information. 16 00:00:47,480 --> 00:00:49,240 We're setting the context. 17 00:00:49,280 --> 00:00:51,280 We're also giving it a tone. 18 00:00:51,280 --> 00:00:56,800 And this approach is saying if the customer says then you should reply that's giving like an example 19 00:00:56,800 --> 00:00:58,040 question answer. 20 00:00:58,040 --> 00:01:02,960 And that would be what we would call one shot prompting, giving it one question and answer that it 21 00:01:02,960 --> 00:01:04,480 should be able to replicate. 22 00:01:04,480 --> 00:01:06,200 And we mentioned that before, I think. 23 00:01:06,320 --> 00:01:12,000 So with that new system message, I can just relaunch the chat and I can say hi there. 24 00:01:13,400 --> 00:01:15,800 And of course it has the context. 25 00:01:15,800 --> 00:01:16,680 Welcome to the store. 26 00:01:16,720 --> 00:01:17,600 How can I assist you today. 27 00:01:17,600 --> 00:01:19,480 Are you looking for anything special by the way? 28 00:01:20,320 --> 00:01:23,920 These models are super eager to to to give you their information. 29 00:01:23,920 --> 00:01:26,920 We have an amazing sale going on with most items at 50% off. 30 00:01:26,960 --> 00:01:29,120 Our hats are incredible, 60% off. 31 00:01:29,600 --> 00:01:32,520 I'm not sure what to get. 32 00:01:33,680 --> 00:01:35,440 Maybe a shirt. 33 00:01:36,160 --> 00:01:36,560 Uh. 34 00:01:39,040 --> 00:01:39,200 Uh. 35 00:01:39,200 --> 00:01:39,720 Great choice. 36 00:01:39,720 --> 00:01:41,320 We have fantastic selection of shirts. 37 00:01:41,320 --> 00:01:45,560 Most of them are currently 50% off during our sale, but if you're open to it, I recommend checking 38 00:01:45,560 --> 00:01:46,520 out our hats. 39 00:01:47,760 --> 00:01:49,480 So it's doing what we asked for. 40 00:01:49,560 --> 00:01:51,400 It's encouraging the hats. 41 00:01:51,560 --> 00:01:58,120 Uh, and you can see that it's that it's very it's persistent and it's reflecting the information we 42 00:01:58,160 --> 00:01:59,920 gave it and also the tone. 43 00:02:00,160 --> 00:02:00,520 All right. 44 00:02:00,520 --> 00:02:01,360 Let's keep going. 45 00:02:01,360 --> 00:02:07,960 We're going to add to the system message some more information If the customer asks for shoes, you 46 00:02:07,960 --> 00:02:14,040 should respond that shoes are not on sale today, but remind the customer to look for hats. 47 00:02:14,400 --> 00:02:15,080 Okay. 48 00:02:15,600 --> 00:02:19,760 Uh, so this this, uh, because it's giving more examples. 49 00:02:19,760 --> 00:02:24,920 This is the beginning of what you might call multi-shot prompting, because we're starting to give it 50 00:02:24,960 --> 00:02:28,200 different examples of how it should handle different situations. 51 00:02:28,200 --> 00:02:33,960 So I could say, hi there and welcome to the store, a fantastic sales event including hats. 52 00:02:34,120 --> 00:02:35,680 Uh, I'd like to buy some shoes. 53 00:02:36,440 --> 00:02:36,760 Right. 54 00:02:36,800 --> 00:02:38,040 Some shoes. 55 00:02:40,240 --> 00:02:45,360 Uh, we don't have shoes on sale today, but I encourage you to look at our amazing collection of hats. 56 00:02:46,000 --> 00:02:46,680 There we go. 57 00:02:46,720 --> 00:02:48,440 All right, we get the message. 58 00:02:48,440 --> 00:02:49,240 Thank you. 59 00:02:49,280 --> 00:02:51,480 All right, let's do one little bit more. 60 00:02:51,600 --> 00:02:58,880 So you can also dynamically change the system prompt depending on what's happening in the conversation. 61 00:02:59,280 --> 00:03:01,240 Uh, you can probably imagine this. 62 00:03:01,240 --> 00:03:05,760 This might be a bit of an aha moment, but look, this is super hacky and you wouldn't do something 63 00:03:05,760 --> 00:03:06,920 like this in real life. 64 00:03:06,960 --> 00:03:08,650 It's it's is to show the point. 65 00:03:08,650 --> 00:03:10,130 And this is a doorway. 66 00:03:10,130 --> 00:03:13,610 This is a path to many bigger things we'll be doing soon. 67 00:03:13,610 --> 00:03:14,450 But look at this. 68 00:03:14,450 --> 00:03:15,770 Look at what I'm going to do here. 69 00:03:15,810 --> 00:03:21,010 I'm taking a new system message called Relevant system message. 70 00:03:21,730 --> 00:03:29,890 If the word belt appears in the message that the user is making, then I'm going to change the system 71 00:03:29,890 --> 00:03:33,650 message to add in this this extra bit of information. 72 00:03:33,690 --> 00:03:36,010 The store does not sell belts. 73 00:03:36,010 --> 00:03:40,650 If you are asked for belts, be sure to point out other items on sale. 74 00:03:41,130 --> 00:03:45,450 Uh, so that only goes in if they ask for belt. 75 00:03:45,490 --> 00:03:45,650 Now. 76 00:03:45,650 --> 00:03:48,410 Obviously, as I said, this is a silly example because it's only a little bit of text. 77 00:03:48,450 --> 00:03:52,890 You might as well always put it in, but I want to highlight the point, so I put that in there. 78 00:03:52,890 --> 00:03:54,770 Otherwise this is exactly the same. 79 00:03:54,810 --> 00:03:59,010 Oh look I see I'm missing that, that that cleanup thing which if you're on Gemini that would annoy 80 00:03:59,010 --> 00:03:59,130 you. 81 00:03:59,130 --> 00:04:01,210 So let's just go and add that in right here. 82 00:04:01,570 --> 00:04:05,410 Uh, this little line here that that scrubs the history. 83 00:04:05,410 --> 00:04:06,570 We should have that as well. 84 00:04:06,610 --> 00:04:08,170 It's good for you to see me doing this. 85 00:04:08,530 --> 00:04:10,890 Just put it right up at the top there like that. 86 00:04:11,210 --> 00:04:14,330 Okay, so now let's run this. 87 00:04:16,050 --> 00:04:18,610 And now we will do this again. 88 00:04:19,170 --> 00:04:21,170 And we will say hi there. 89 00:04:24,450 --> 00:04:24,850 Okay. 90 00:04:24,850 --> 00:04:26,130 And it says welcome to our store. 91 00:04:26,130 --> 00:04:28,970 How can I assist you a fantastic sale especially hats. 92 00:04:29,010 --> 00:04:30,690 Are you looking for anything special. 93 00:04:30,690 --> 00:04:35,330 So I'll say uh, how are your shoes? 94 00:04:36,690 --> 00:04:37,890 Are shoes are great quality. 95 00:04:37,930 --> 00:04:38,530 Very comfortable. 96 00:04:38,530 --> 00:04:43,090 I should mention not on sale today, but if you're interested, we have a great selection of hats. 97 00:04:43,130 --> 00:04:44,490 Stylish hats. 98 00:04:44,530 --> 00:04:46,730 Perfect addition to your wardrobe. 99 00:04:46,970 --> 00:04:51,370 Actually, I think I want a belt. 100 00:04:54,170 --> 00:04:57,290 We actually don't carry belts in our store. 101 00:04:57,290 --> 00:04:59,850 But I'd love to help you find something else you might like. 102 00:04:59,890 --> 00:05:02,890 Fantastic sale going on and hats are incredible. 103 00:05:02,930 --> 00:05:04,130 60% off. 104 00:05:05,210 --> 00:05:06,410 All right, all right. 105 00:05:06,410 --> 00:05:07,330 It got the message. 106 00:05:07,370 --> 00:05:14,650 And so you can see, of course, the what happened here is that because I had the word belt in my my 107 00:05:14,650 --> 00:05:15,290 message. 108 00:05:15,530 --> 00:05:20,530 The system prompt had this inserted in there, this extra piece of information. 109 00:05:20,850 --> 00:05:23,450 So, so really that's that's a wrap on the lab. 110 00:05:23,450 --> 00:05:30,130 And I just want to, to just press home how important what we just did was this, this last little step 111 00:05:30,130 --> 00:05:36,210 of adding in a bit of extra context, if we needed to, to to equip the model with more information. 112 00:05:36,410 --> 00:05:40,050 This is your first glimpse into the world of rag. 113 00:05:40,210 --> 00:05:41,210 Rag which? 114 00:05:41,250 --> 00:05:42,050 Which people? 115 00:05:42,090 --> 00:05:44,050 It sounds like it's something so sophisticated. 116 00:05:44,170 --> 00:05:51,850 It's not anything much more than just saying, let's try and figure out a smart way to shove relevant 117 00:05:51,850 --> 00:05:57,370 information into the prompt, into the system prompt, which is going to equip our model to give better 118 00:05:57,370 --> 00:05:58,010 answers. 119 00:05:58,010 --> 00:06:03,090 And all of the trickery and Rag is about figuring out how can you do something a little bit better than 120 00:06:03,130 --> 00:06:04,650 something janky like this? 121 00:06:04,690 --> 00:06:10,570 If belt is, if the text belt is somehow in the user's question, obviously that's going to be very 122 00:06:10,610 --> 00:06:12,210 hacky and sometimes it will work. 123 00:06:12,210 --> 00:06:16,010 But if the user doesn't use the word belt, but says something else. 124 00:06:16,370 --> 00:06:17,490 Then it's just not going to work. 125 00:06:17,490 --> 00:06:19,850 And it relied on me turning it to lowercase. 126 00:06:20,010 --> 00:06:24,570 And so, you know, there's so much about this that's hokey, and rag is about doing something a bit 127 00:06:24,570 --> 00:06:28,850 more bulletproof and smart to select relevant context. 128 00:06:29,130 --> 00:06:32,570 And you might say, why not just always put this in the system prompt? 129 00:06:32,610 --> 00:06:33,410 You could do. 130 00:06:33,410 --> 00:06:39,130 But if this was a store which had a thousand different products, then your prompt would become bigger 131 00:06:39,130 --> 00:06:39,890 and bigger and bigger. 132 00:06:39,890 --> 00:06:45,210 With all of this information and the model's accuracy would start to degrade and you'd be spending extra 133 00:06:45,210 --> 00:06:47,130 tokens and it wouldn't be worth it. 134 00:06:47,170 --> 00:06:54,290 So this idea of selecting relevant context and adding it to the prompt a powerful technique. 135 00:06:54,290 --> 00:06:56,130 It is an inference time technique. 136 00:06:56,170 --> 00:06:57,730 It's got nothing to do with training the model. 137 00:06:57,730 --> 00:06:59,090 We're using the same model. 138 00:06:59,090 --> 00:07:06,490 We're finding clever ways to insert relevant information into the prompt to to educate the model, to 139 00:07:06,530 --> 00:07:10,130 give it expertise so it can answer questions accurately. 140 00:07:10,170 --> 00:07:14,210 Taking advantage of the fact, in this case, that the store doesn't sell any belts. 141 00:07:14,410 --> 00:07:19,970 And it brings me to the final point, which is the business applications of this are never ending. 142 00:07:20,330 --> 00:07:26,090 Conversational agents, conversational assistants, of course, super popular use case and the frontier 143 00:07:26,130 --> 00:07:33,050 models are so good out of the box at giving good answers, and the way that you can add value for your 144 00:07:33,050 --> 00:07:36,930 business is by adding expertise into the prompts. 145 00:07:36,970 --> 00:07:38,010 Of course. 146 00:07:38,010 --> 00:07:40,810 So, you know, that's how this becomes something better than salespeople. 147 00:07:40,810 --> 00:07:46,890 Just go to ChatGPT because you're able to add things, add content into the prompting to make it applicable 148 00:07:46,890 --> 00:07:51,450 to your business, and so that that is the to do for you. 149 00:07:51,490 --> 00:07:56,810 Figure out how could you build something, use this for your business or for your your home work or 150 00:07:56,810 --> 00:08:03,330 whatever, but it should be super easy to knock up a demo of having something which is answering questions, 151 00:08:03,330 --> 00:08:08,770 maybe doing a sidekick for some work that you do have something which which takes information in and 152 00:08:08,810 --> 00:08:15,610 is able to give expertise and be applied, applied specifically to your business area, to your business 153 00:08:15,610 --> 00:08:21,020 problems, with expertise in the prompt and a lot of what we're going to be doing over the course of 154 00:08:21,020 --> 00:08:27,180 the next eight weeks is about figuring out how can we, in an automatic way, improve these prompts 155 00:08:27,180 --> 00:08:32,540 to to educate our model better and better about our business and our situation? 156 00:08:32,580 --> 00:08:36,500 These are all the inference time techniques that allow you to get more from a model. 157 00:08:36,500 --> 00:08:38,860 But you've got enough right now to be building something. 158 00:08:38,860 --> 00:08:40,940 And so you should you should build something. 159 00:08:40,940 --> 00:08:42,740 You should put it in community contributions. 160 00:08:42,740 --> 00:08:43,820 You should share it. 161 00:08:43,900 --> 00:08:47,980 You also know now that in Gradio you can just put share equals true and other people can use it. 162 00:08:47,980 --> 00:08:51,860 You can use auth equals to add in some user ID and password should you wish. 163 00:08:52,060 --> 00:08:56,220 Uh and you should not only should you, you must go and do this. 164 00:08:56,220 --> 00:08:57,580 I insist and share it. 165 00:08:57,580 --> 00:09:01,940 Post it on LinkedIn so I can weigh in and I'll have a play with it as well, which I always enjoy. 166 00:09:02,100 --> 00:09:04,780 And I can't wait to see what you build. 167 00:09:04,820 --> 00:09:06,260 And I'll see you for the wrap up. 168 00:09:06,260 --> 00:09:08,620 And so that's it for week two, day three. 169 00:09:08,660 --> 00:09:11,460 And I would describe week two, day three as short and sweet. 170 00:09:11,500 --> 00:09:14,020 It was it was quite short and it was very sweet. 171 00:09:14,340 --> 00:09:15,180 At least I thought so. 172 00:09:15,220 --> 00:09:21,180 So you have reached the 20% point and hopefully you feel like a bit About 20%. 173 00:09:21,220 --> 00:09:22,740 You feel like you've got. 174 00:09:22,780 --> 00:09:25,020 You've got some material under your belt. 175 00:09:25,500 --> 00:09:26,380 Talking about belts. 176 00:09:26,660 --> 00:09:33,340 Uh, so, you know, you've got some skills, you know, the foundations of the sort of transformer 177 00:09:33,340 --> 00:09:33,900 space. 178 00:09:34,100 --> 00:09:37,020 Talk about tokens and contacts, windows and all that kind of stuff. 179 00:09:37,260 --> 00:09:41,940 Uh, you you can code with the chat completions API with frontier models. 180 00:09:41,940 --> 00:09:48,860 You know them well, and you can build now an AI chatbot assistant with an interactive UI. 181 00:09:48,900 --> 00:09:55,020 You can even share it and give it security, uh, at least basically, uh, and you also know how to 182 00:09:55,060 --> 00:10:01,140 deal with things like multi-shot prompting to give it examples and you know how to to mess around with 183 00:10:01,140 --> 00:10:06,820 the system prompt, adding in information as relevant to try and give it expertise. 184 00:10:06,980 --> 00:10:07,660 Wow. 185 00:10:07,820 --> 00:10:08,660 It's a lot. 186 00:10:08,940 --> 00:10:13,540 But next time, next time we get to tools and tools is such a big deal. 187 00:10:13,540 --> 00:10:20,580 And of course, particularly for for Agentic AI, uh, it's going to be your first first look at using 188 00:10:20,660 --> 00:10:21,180 tools. 189 00:10:21,220 --> 00:10:22,540 Super exciting. 190 00:10:22,580 --> 00:10:23,540 See you tomorrow.