so suj have you looked in your LinkedIn profile lately and noticed there are a ton of job openings for prompt Engineers absolutely and that's why today we're going to do a deep dive on what that is and but first to give a little context let's talk about what large language models are used to do for a review of course everyone is familiar with chat Bots and that's seen that all the time it's also used for well using summaries for example another common use case or information retrieve those are three different cases but for our viewers
could you explain how that applies in prompt engineering sure prompt engineering is very vital in communicating effectively with large language models what does it mean it is designing coming up with the proper questions to get the responses you looking for from the large language model because you want to avoid hallucination right hallucinations are where you get essentially false results out of a large language models and that's because because the uh large language models are predominantly trained on the internet data and there could be conflicting data conflicting information and so on great okay I got that
so we're going to look at this from four different approaches so let's get straight to it yep we're going to look at the first approach which is rag or retrieval augmented generation we've had videos about this already on the channel so I have kind of a basic understanding of it where you take domain specific knowledge and add it to your model but how does that actually work behind the scenes could you explain that to me absolutely so the larger language models as you know are trained on the internet data they are not aware of your
domain specific knowledge base content at all so when you are quering the large language models you want to bring awareness of your knowledge base to the large language model So when you say knowledge base here you're referring to something that might be specific to my industry specific to my company which I'm going to then be applied to the model absolutely and so as that work again so to make this uh bring this awareness to the large language models we have to have two components one is the retriever component which brings the context of your domain
knowledge base to the generator part of the large language model and when they work together and when you ask questions to the large language model it is now responding to your questions based on the domain specificity of your content okay I think I got it now this retriever that could be really as simple as a database search right exactly it can be a dector database okay I got that um but could you first kind of give me a quick example of how you seen that applied in an industry absolutely let's take the example of a
financial uh information for a company right if you were to directly asking a question through the large language model about the total earnings of a company for a specific year it's going to go through its learning and the internet data and come up with a number that may not be accurate right uh so for example the annual earnings it could come back with $19.5 billion and which may be totally incorrect whereas if you want to get the accurate responses then you bring the attention to the domain knowledge base and ask the same question then the
large language model is going to refer to your knowledge base to bring that answer and this time it will be accurate say for example $5.4 billion I see because this is a trust and source that it can then integrate in with this larger model correct okay so now we're on to the second approach to prompt engineering C or Chain of Thought and I I sometimes think of this as the old saying explain it to me like I'm an eight-year-old but could you give me more a practical explanation what that really means absolutely I think uh
the large language models like in 8-year-old also need guidance on how to arrive at those responses right and before I jump to the um Chain of Thought approach I want to um recommend something right anytime you are working with the large language models consider two things the number one is the rag approach content grounding content ground your uh large language model right and then take the approach of promting it guiding the model through the prompts to get the responses that you need and cart belongs in that category as well these other three absolutely so let's
talk about Chain of Thought right Chain of Thought is all about taking a bigger task of arriving at a response breaking it down into multiple sections and then combining the results of all those multiple sections and coming up with the final answer so instead of asking a large language model what is the um total earnings of a company in 2022 which it will give you just a BL blur of a number like $5.4 million you can actually as a large language model give me the uh total earnings of a company in 2022 for software for
hardware and uh for uh Consulting say for example I see so you're asking to be more precise with the idea that you'll be able to get individual results that will ultimately combine combine it I see so for example you cited we'll just make up some numbers if I had five and then continue the rest of and three for example and the final answer will be 5 + 2 + 3 that will be the output but the large language model is now arriving at this number uh through reasoning and through explainability the was these was three
separate queries essentially three separate problems so the way I tell the large language model is I give the problem and I explain it on how I will break down the problem so for example I say what what is the uh total earnings of a company and the if the total earnings of a company for software is five for Hardware it is two for Consulting it is three then the total earnings is 5 plus 2 plus 3 let me see if I can net that out to make sure I got it so in rag we were
talking about being able to essentially improve based on domain knowledge but then to improve on the results that that generates we then apply this technique the explain it to an 8-year-old technique which then makes the result even better mhm okay that was Chain of Thought which as I understand is a few shot prompt technique where you basically provide some examples to improve the end result and I think the react is kind of the same genre but it's a little bit different could you explain to me the difference absolutely so react is also a few short
pting technique U but it is different than the Chain of Thought in Chain of Thought you are going breaking down the steps of arriving at the response right so you were reasoning through the steps and arriving at the response whereas react is goes one step further it not only re reasoning with that but acting based off of what else is necessary to arrive at the response so this data though is coming from different sources we weren't talking about that in the latter case with k f and they are so they are for example you have
a situation where you have your content the the domain content in your private database knowledge base right but you are asking a promp where you question is demanding responses that are not already available in your knowledge base then the react approach has the ability to actually go into a private a public knowledge base and gather both the information and arrive at the response so the action part of the react is its ability to go to the external resources to gain additional information to rual responses I got it I got it but there's one thing that's
confused me just a teeny bit is that in rag that looks awfully similar but they're not the same where's the difference here so the difference is uh they both are using the private uh databases right knowledge basis but in large language models I want you to think about two steps right one is content grounding that's what rag is doing it is making you large language model aware of your main content where react is different is it has the ability to go to the public resources public content and knowled knowledge basis to bring additional information to
complete the task okay uh before we wrap can you give me an example of react absolutely so let's go back to the financial example you were looking at in the previous uh patterns we were looking at the total earnings of a company for a specific year now supposing you come back with a prompt where you were ask asking for the total learnings of 2010 and 2022 right your 2022 is information is here in your private database knowledge base but 2010 information is not there for example it's over here in the public one exactly so the
large language model in the react approach now take takes the extern takes to the external resources to get that information for 2010 and then brings both of them and does the observation I see so that's going to produce a result that takes into consideration this whereas before it might have produced essentially a hallucination hallucination and a couple of more things right the react gives you the results in a three-step process right when you are asking the prompts in a react mode you have to first of all split that prompt into three steps one is the
thought right what are you looking for and the second one is action what are you getting uh from where right and the third one finally is the observation that is the summary of the action that is taking place so for example thought one will be retrieve the total earnings for 2022 right and the thought so action one will be it will actually go to the knowledge base to retrieve 2022 and observation will be 2022 value now thought two is ret the value for 2010 from a an external knowledge base and have that value there and
observation two will have that value and the part three will be comparing them to arrive at which is a better Total Learning for you I think I've got it that's great we only have one more to go if you really want to impress your colleagues you want to learn about this next one which is directional stimulus prompting or DSP different from the other ones and how so DSP is a fun way uh and a brand new one that I want to introduce to the audience of uh uh making the large language models give specific information
giving it a direction to give specific information from the task so for example you ask a question and say for example what is the um annual earnings of a company uh but then you want don't want the final number but you want specific details about learnings for say software or for Consulting so you give a hint and say software and Consulting and the large language model first of all we'll get the earnings and then from that extract specific values for software and Consulting this kind of reminds me of the game where you're trying to get
someone to draw a picture and what do you do you provide a hint and in effect this provides you a better result in the same fashion absolutely so it is a very simple technique but it works very very well when you are looking for specific values from the task so try it out well thanks Su I Now understand what DSP is but could you kind of net out how do you combine these different techniques um you you should always start with rag to bring Focus to your domain content but you can also combine cot and
react you can also combine Rag and DSP to get that cumulative effort uh effect excellent okay well thank you very much I hope you come back for another episode in promptu absolutely thank you Dan thank you for watching before you leave please click subscribe and like
