Loading...
So, the stove's in your kitchen. You can finally start cooking. Last section felt a little bit boring, a little too much blah blah blah, but you have the oven, the stove, the big object ready to cook your AI. What are you going to cook on this stove? You're going to cook functions, semantic functions. What kind of cooking are you going to do? Business style stuff. We're going to cook up. SWOTs, strength, weakness, opportunity, threats, boring business stuff, and make it sing on your stovetop with AI. Let's see it happen. Semantic functions, functions made out of words, so-called prompts, and basically do things that you might not expect are possible, which is scale the kind of AI you're trying to build. Not just scale the number of people that can use it but scale the kind of system you can build with it. There are two kinds of functions in the semantic kernel world. One is a native function, classic computer code, you type it in, and it runs, or semantic functions composed of text that is parameterized. there's some kind of configuration attached to it, and it produces output as well as a native function, but you'll find that these semantic functions are usually powerful. Reminder what's a SWOT. A SWOT is strength, weaknesses, opportunities and threats. It's a good matrix to use when helping a business of any scale understand what it should be worried about. Okay we're going to get a kernel ready, and I'm going to grab from the semantic kernel as SK. You'll see me type this a lot, and you'll get bored very quickly. So, let me just copy and paste a bunch of stuff, and we're going to run that. The kernel is now ready. Next up we're going to do is we're going to make our first semantic function. Okay, let's do this. So, what we're going to do is we're going to make a prompt, a templated prompt. Let's read this for a second. It's a Python string so it's a multi-line string. It has a certain input, and the prompt is to summarize whatever is above in less than 140 characters. We're then going to take a detour to make a semantic function that takes this prompt template. There are some other parameters. That's boilerplate. Let me just add that for a second. There we go. I'm adding a description. I am telling how many tokens to use, setting the temperature, and giving the top p to describe the range of words to be used in the response, the completion response. And just a reminder, because this isn't going to run, and you won't see anything happen. Let me just remind myself that what it's doing is it's making a semantic function for summarization and now it's registered. So next up, if you notice this value input, I want to be able to attach some value to that. And I'm going to take text from a TED talk from Andrew Ng about a pizza shop owner who could use some help from AI, he thinks. And I'm going to then compute the following. I'm gonna generate a result by having the kernel do a run async. This is now adding the summary function into the pipeline. And I'm gonna give it at the head of the pipeline, SK input. And let's print this out in a pretty way. I'm gonna use my display markdown thing. And let's see how it does. Let's see how it's doing. It's running, great. It took this information, and it summarized it. What did it do? It basically took this summary function, it stuffed this input into the summary function, and then it generated, Sparkle, this short summary of that long text. You know, long text, nothing's wrong with it, long form is good. Okay, how are you feeling so far? Kind of cool, right? Let me just tell you there's another shorthand way of calling these different functions without using the kernel. In this case, I'm just going to call the function directly, just kind of like calling a procedure. You'll see this used once in a while by people. They want to show that you don't have to use it with an explicit kernel. And there it is. It did it the same way, but this is much more compact. It doesn't using this await kernel runs async, but does the same thing. Same result. See that? Okay. High feeling. Okay, we're going to go a little harder now in the world of native functions. Native functions. Native functions are not semantic functions, and they require a lot of so-called syntactic sugar meaning they they're wrapped in a bunch of things that I could type but pardon me for typing fast meaning cut and paste let's pull these things in we're going to make an exotic language plugin it's going to convert anything into. Pig Latin as you can see what it's doing is I define a semantic kernel function I give it a description I give it a name. I tell it what is this input variable, you know this input like what do I do with the input what is it. What do I do? Do this. What is it? It's this kind of text. It takes in Pig Latin, sentence comes in, it walks over the words, and of course it processes it. So here I'm going to say exotic language plugin, I'm going to register it in the kernel. I'm going to register this plugin. See that? A lot of syntax, right? But wait for it. This is the exciting part. I'm going to define the function. This is kind of, what's it called again? This is not going to feel awesome but know this is a big deal. You're taking a native function, and you're wrapping it in a format that Semantic Kernel can use. In the same way that Semantic Kernel can process semantic functions, it can process native functions. And next up, what we want to do is we want to run a pipeline of functions, The summary function and the Pig Latin function with the input. So, the Andrew Ing long form will go into summary function. It'll be fed as an output into the Pig Latin function, and guess what's going to happen? I'll assume you'll guess what's going to happen, but yeah. It's basically that sentence before, AI can blah, blah, blah, turned into Pig Latin, okay. Clap, clap, clap, applause, AI, but now you can see that it's using a native function, and a semantic function so you can interchange them pretty cool and you know, and all this like techno stuff you can get lost, I get lost pretty quickly myself, because it's so powerful. But end of the day, all you're doing is writing code to solve a problem, so the TED talk that Andrew Ing gave about how to help small businesses in particular a pizza shop owner struck a chord with me, and for that reason I have built this entire lesson for you to not just write code, chord with me, and for that reason I have built this entire lesson for you to not just write code, but think in business terms. So, let's sort of take off by, first of all, thinking of the SWOT instrument. To create a SWOT, you have to create a list of questions. And this is a fairly good list of questions that are composed to figure out, like a small business owner, like what are their strengths, weaks, opportunities, and threats? And if you organize these in a SWOT chart, the way they look is this. Let me show you this. I love a two-by-two, don't you? So, these are the strengths. Look at them quickly. Unique garlic pizza recipe, owner trained in Sicily, weaknesses, don't have calzones, had a flood in the area, the seating areas are in disrepair, opportunities, they don't do catering or cater to the local tech startup community. There's also an annual food fair. This is a way they could actually make new revenue. And there's threats. There are cheaper pizza shops, there's gonna be street construction, the cost of cheese is going up. And so, this is a snapshot, a swap. So, what does this have to do with large language models and semantic kernel? Good question. Well, first off, remember, it is a way to solve AI problems, semantic kernel, but it doesn't matter if you don't have a good problem. Now, let's make a kernel again. I'm gonna make a kernel. This is very familiar to you. And again, I just put print statements in to just remind myself. Made a kernel. I made a kernel. Well, and now what I want to do is I want to take these SWOT responses, analyses, and. I want to convert them to a different domain. Why would I want to do that? Well, because as they say, the mountain says you can, and AI gives you the ability to climb mountains so quickly. It's unbelievable. So, let me show how that works. So, what I've done here is I've taken the SWAT interview and I'm making a semantic function that converts the analysis into a different business domain. Kind of weird, huh? And what I'm going to do is I'm going to take this interview text, I'm going to ask it to apply it to the construction management domain. And I'm going to run it. And what happens is it takes questions that were geared for a pizza shop, and now I could give them to someone in construction management. Which sounds kind of science fiction-y, but here, it's changed all the pizza-esque things into construction methods, materials, budget overruns, do you see the zoning regulations? And so, it basically like shifted the context of the text. And that is something you cannot write in a native function, but in a semantic function, it's unusually easy to do. Let's throw in another, cause we like to chain things together. Let me show you another type of semantic function. This is a rewrite. Rewrite the text above to be understand by a blank level. So, what we're doing is we're making a new function to be able to change the reading level of whatever comes into it. And what I'm gonna do is I'm gonna redo the shift domain function, shift the questions into another domain, construction management. I'm going to flow that in the pipeline to changing the reading level of whatever comes into it to the level of a child. It's a lot easier to read the good things, the bad things, the good chances, the bad chances and there you have it. And I'd like you to change any of these parameters and see what happens. And I think you'll get pretty excited because I don't think you've ever written programs like this. If you're an LLM person or ML person in general, this is no surprise to you. But for someone working primarily in the domain of problems, it's extraordinary. And you want to remember that we are using this, the right hand, I talked about the semantic completion ability, fill in the blanks. This is the magic that we're experiencing right now. But there's this other type of AI capability that we're using out there called semantic similarity. I haven't touched upon that yet, so just be ready for when you notice that you can use both hands in the equation.
Developing Business Applications using LLMs via Semantic Kernel
- Introduction
- Semantic Kernel is Like Your AI Cooking Kitchen
- Cooking Up Flavorful SWOTs with the Kernel
- Organizing The Tools You Make for Later Reuse
- Frozen Dinner The Design Thinking Meal
- Dont Forget to Save the Generated Dripping or The Gravy
- A Kitchen That Responds to Your I'm Hungry is More Than Feasible
- There's a Fully-Outfitted Professional-Grade Kitchen Ready For You
- Conclusion