Up next in 10
Unleashing the Power of Microsoft Semantic Kernel in API Communication - A Deep Dive into the GPT-4
5K views
Jan 29, 2024
Join this live session with Stephen Simon ft. Daniel Costea for the next episode of A Dash of .NET - live show on January 25, at 09:30 PM (PST). Microsoft Semantic Kernel (SK) is a new technology that enables the integration of AI Large Language Models (LLMs) with conventional programming languages like C#, Python, and Java. SK brings together several key components to provide planning and execution capabilities. These components include a robust kernel that provides the foundation for all other components, plugins (formerly known as skills) for performing specific tasks, connectors for interfacing with external systems, memories for storing information about past events, steps for defining individual actions, and pipelines for organizing complex multi-stage plans. What about if we go beyond those clicks and build a semantic interface (code in C#) for an existing API (controlling a Car Robot)? GUEST SPEAKER Daniel Costea is a senior software engineer with more than two decades of experience and a passion for C# since its inception in 2001. He has been a frequent international speaker in Europe and the US in the past 6 years, and has organized community meetings and workshops. As a trainer, Daniel has delivered over 1200 hours of training in .NET web technologies to more than 700 students. He is also a published author of articles on .NET technologies, and was recognized as a Microsoft MVP in developer technologies in 2019. 👉 Daniel Coste Twitter: https://twitter.com/dfcostea 📺 CSharp TV - Dev Streaming Destination http://csharp.tv C# Corner - Community of Software and Data Developers https://www.c-sharpcorner.com #azure #CSharpTV #csharpcorner #dotNet
View Video Transcript
0:00
Hi everyone and welcome back to this Dash of .NET show
0:03
I'm really, really excited for this episode because we're going to learn something really new
0:07
which I have not seen anyone talking about recently in the community and to deliver this great session today
0:14
we have Daniel who is a Microsoft MVP and a Senior Software Engineer
0:18
We have hosted him quite a few times in last three, four years and I'm so happy to have him back
0:24
So without any further ado, let's bring back the rockstar Daniel. Hi Daniel, welcome to the Dash of .NET show
0:34
Thank you very much. I'm very happy to be back. Yeah, Daniel. I'll compliment you straight up
0:40
You look more fine than you did last year. You have changed your look
0:44
It was fun day. Yeah, that makes sense Daniel. So Daniel, what are you going to talk about today
0:52
Well, I'm very excited about this topic. I mean, I will talk about Microsoft Semantic Kernel
0:59
I will jump in a bit and explain what it is. But nevertheless, I have a chance to talk about something very
1:06
new because we just had the chance to see the final release last month
1:12
Yeah, that was really cool to have it released this month and I've seen a few people writing blogs and making videos around it
1:20
I'm so happy that you're going to talk about today. So Daniel, I'm not going to take much of your time
1:24
Please feel free to go ahead and share your screen and then we can get started
1:29
Very well. Let me share the screen. And everyone who's watching, you can find Daniel's social handles
1:35
in the video description. If you want to follow him on Twitter, on LinkedIn, you can get all the details from there and reach out to him just in case
1:42
if you need any help. I'm sure Daniel is very helpful. I mean, that's what he does, right
1:48
He's a community guy. I see your screen. Yeah, that's your presentation
1:54
I'm going to add your presentation to the stream now. Everybody else can see your screen and it's all you now
2:00
Fantastic. All good? OK, so firstly, why Semantic Kernel
2:11
We had the chance to see some other frameworks. For example, there is a library for consuming OpenAI LLM models
2:24
but that was more like a client. And now we are going to talk about an orchestrator
2:34
Semantic Kernel is more than a client. It's not only sending prompts back and forth
2:40
but it's about using the power of AI to orchestrate some more complex scenario
2:50
So this is Semantic Kernel. And I will let you know in a bit what it is
3:00
So it's an open source. That means you can contribute. You can influence what is developing there
3:07
And I have to tell you that in the last month, there were a lot of changes
3:14
because the last month was just released. The final release, 1.0.1, took place a few weeks ago
3:24
And until then, for about half a year, there were a lot of changes
3:30
And why was that? It was because OpenAI had a lot of changes
3:36
They added function calling. They added assistance, a lot of features. And these had a direct impact on the development of this library
3:51
But maybe the most important part, it's that this library is for Toddler developers
4:00
We had a chance maybe to see a link chain, but that was limited to Python and JavaScript
4:10
if I'm not wrong. But now with Semantic Kernel, C Sharp developers, along with Python and along with Java
4:19
had the chance to use this power of large language models. This library is able to consume services directly from OpenAI along with Azure OpenAI
4:36
but not limited to them because it's able to consume models from HuggingFace
4:43
And maybe, again, it's not very... Actually, I didn't check recently, but in the past, I had the chance to see that this library
4:58
being abstract, is able to consume on-premises models like Lama family models
5:06
And that's fantastic because you have the chance to run everything on-premises in a silo
5:14
I don't know, in a sandbox and play with it. So, as I said before, this is an orchestrator
5:23
And how is this completed? It's by using some containers for semantic functions and native functions
5:36
But what are these? We will see in a bit. Firstly, I would like to tell you that in order to build advanced prompts, we have here a syntax
5:54
And this syntax is able to work with some placeholders for using some variables or even
6:03
functions with arguments. And that's very good because we don't want our prompts to be static text
6:13
We are happy to make them dynamic, injecting some data in these prompts on the fly
6:24
So, again, what are these plugins? Containers, where we can put and we can group together functions which are for the same purpose
6:39
for example. We can even mix native functions with prompt functions. But let's say we have a group of functions which are processing text
6:53
Yeah, we would like to put them together like, I don't know, string functions or strings plugin
6:59
whatever the name you like. And in the same time, we can use these plugins to keep some native functions
7:11
Let me get into some details. What are each of them? Semantic functions are text
7:20
Text in, text out. It's a concept we had a chance to hear before
7:25
But it's like natural language functions, which are, of course, what large language models are using
7:37
But we are not limited to that because we know these large language models were trained
7:45
using some data which are not very current. So, in order to work with some real-time data, we need to do some, for example, some search
8:02
on Internet. And for that, we need some native functions to call some endpoints
8:09
Or let me tell you about some other scenario. For example, you have an API
8:14
And this API has some specific endpoints. We would like to be able to call them using HTTP client, for example
8:25
But there are a lot of out-of-the-box native libraries, plugins, sorry, native plugins
8:36
included in Semantic Kernel. And we can develop our own, which I will demo in a bit in this talk
8:47
So, as soon as we have these plugins, we can use them or we can export them in ChargPT
8:58
if we like. Or Bing chat or Microsoft 365. Let me show you a diagram
9:07
So, we have different plugins. And we can mix them together and solve some use cases
9:22
We are going to keep these plugins in our kernel. Kernel is a concept of this Semantic Kernel, which is the glue where we put together
9:37
plugins and connectors and some other settings. And, of course, we are able to call these functions
9:52
Of course, we can create the plugins in a programmatic way. But we are not limited to this
10:00
We can have some dedicated way of working with these plugins by using some special files
10:09
And these are escape prompt.txt, which is, in fact, a JSON file, which keeps
10:18
information about the prompt and the variables used for this prompt. For example, when we try to get a response from a large language model, if we want to be
10:39
creative because we are, I don't know, composing a poem or something or a story, we want this
10:46
LLM, large language model, to be very creative. Therefore, we would like to put the temperature higher
10:55
And if we want something which is precise, like generating Python or C-sharp code
11:04
then we don't need creativity, we need preciseness. So then we will put temperature to zero
11:12
But there are some other parameters and it's interesting to play with
11:16
And as I said before, we can put some information about the input variables
11:25
By default, we have the name input for input variable, but we can have more and we can
11:34
customize the names. So in the right side, we have the prompt
11:42
As we can see, it's an English language, but we are not limited to English
11:47
You can try playing with your native language. Why not? And inside what is not so common are these double accolades, which is our dynamic piece
12:08
of our prompt. For example, the input variables, or we can have a function we are calling
12:17
There is another one here, it's a comment, right? So we have here a prompt which will replace this text with some injected data from comments
12:31
and some injected data from input. Of course, I have used some other tricks in order to emphasize the most important parts
12:41
of this input, but it's not the case to build here ideal prompt, but it's important to make
12:49
it efficient. So having these semantic functions or prompt functions, let's talk about a little bit about
13:00
the native functions. And as we can see, it's a C sharp code
13:07
Similarly, we can have Python or JavaScript. These three languages are not developed equally
13:20
So C sharp and Python are more advanced. I mean, Java may lack some features which are not there yet
13:30
These three languages are not developed equally. So C sharp and Python are more advanced
13:35
I mean, Java may lack some features which are not there yet
13:41
This can be checked by visiting the official repository. And there you can find even a matrix with which features is available for each language
13:59
So let's take a look on this C sharp class. The native functions are easy to identify because they are prefixed by kernel function
14:16
attribute and the description. And the description is very important because the name, actually the text in the description
14:26
has to be concise to not leave space for interpretations. So if you have such class with the kernel functions, keep there only the minimum of
14:45
the functions you are planning to use with your prompt. And be precise and unique with your description because the large language model will try
14:57
to mix and match in order to identify which function is good for your prompt
15:09
So now I think I have to go a little bit in details
15:15
How are these used? For example, I have a robot car and this robot car has a limited functionality
15:31
For example, it can move forward, backward, turn left, turn right and stop
15:40
Basic movements, right? But maybe I want something more complex. For example, walk like a ballerina or do a full turn, a complete turn, 360 degrees
16:04
And these are not very obvious how we'll map with the basic capabilities of our API
16:15
So something, and this is large language model, will have to figure out
16:22
and break down this complex action into basic pieces. For example, if we ask our prompt to do moonwalk dancing, and Michael Jackson walking backwards
16:42
Well, I'm expecting large language model to break down this moonwalk dancing into some
16:52
steps which are going forward more or less. And similarly, for ballerina, I'm expecting to break down this dance in some
17:06
some basic movements like forward, turn left, turn right, you know, capturing a little bit of this dance
17:16
Okay, so I think you got the point. I talk to my API telling a complex move and the large language model
17:28
will break down this complex move into pieces which are very well known by my API
17:42
And let's see how we can do this. Because in order to break down this complex move in pieces, we have to do some tricks
17:55
We have to use some advanced techniques which Semantic Kernel is providing
18:04
And I will talk about function calling and I will talk about planners
18:10
Planners are more, more, it's, it's, they were introduced earlier and they were responsible
18:22
for figuring out complex actions in smaller basic actions or steps. But recently, OpenAI introduced the function calling and these functions are
18:43
some special functions which LLM understands in the large language model. Let me tell you an example
18:54
You can call the large language model with the prompt which asks for the weather in your city
19:07
right now. And of course, the larger models were not trained by current data
19:17
So not a chance to know what's the weather now for your city
19:23
So you can add to your prompt some functions which are able to interrogate some API
19:36
for the weather forecast thing and for the daytime. And having these special functions which are like APIs
19:52
you can send these functions to the large language model. The larger language model will try to identify your needs and it will realize
20:11
okay, in order to, to respond better for this prompt, I have to send back a
20:19
a response to the caller saying I need to call first these functions which were
20:29
sent with along with the prompt. So as I said, the response is coming back with some, this time with some requests
20:38
for these special functions which are running on the client side. They are getting some response, for example, the weather and the daytime
20:51
And then they are sending back, look, look, it's a chat with the large language model
20:56
So the responses are sending back to the large language models. And this cycle is complete when there are no more functions to call because the function
21:11
can call another function. So in the end, this will be complete
21:19
And then the prompt will be augmented with the necessary information. And the larger language model will respond with what you are expecting
21:35
Like what is the weather in your city? And for this daytime in a, in a human format
21:49
Yeah, not like JSON, which is the usual response from an API
21:56
Okay, now there are two flavors on function calling. On function calling, for example, you can call these functions by setting the auto invoke on true
22:11
And this is making our code very simple. As you can see below, it's just a few lines of code
22:21
And let me explain a little. So at some point, we inform our kernel, we are going to use motor plugin here
22:33
and then we are invoking this function. And lastly, and we are iterating through the responses, as I said, is a, is a
22:46
is a chat with large language models until all the functions are solved
22:52
For example, we may expect here to, to have more steps if we break down the ballerina dance
23:03
For example, go left, go right, forward, backward, whatever, and stop, whatever
23:09
So we are expecting here to, to see these steps, which are figured out by this function
23:19
inside the large language model. And of course, having these steps, now the, the, we can do something
23:31
For example, executing these steps on our side. Maybe you have a robot, like I call this presentation, because I started from this idea
23:40
to, to try to talk with the robot API. Yeah, we have, in the end, we have this car
#Business & Productivity Software
#Computer Science
#Intelligent Personal Assistants
#Open Source
#Windows & .NET