[Music] Hey, it's Matt with a repit, and one of the most amazing things about AI and the new developer technologies we're seeing today is that they unlock development not just for experienced software developers, but also for the emerging software creators out there that are just trying to build something cool, um, maybe make their life a bit easier, make their work a bit easier, um, or just learn the fundamentals of programming. So, I'm going to show you a demo today that I built in maybe an hour or two on repet. It's really simple, it's pretty straightforward, and it lets you summarize YouTube transcripts using Gro, which is a free way of interacting with um, some of uh, meta llama models, but also a few other open-source models um, and repet uh, as well as a front end with gradio. All the app really takes is a YouTube URL, and when we click submit, uh, the app is going to fetch uh, that video, get the details, chunk it up, um, and then Su summarize it in segments. So, this is really a basic approach to to

summarization. We're just chunking that video up and summarizing it. Um, but we'll walk through how this works a bit, and then I'll show you how you can use this in your own project. Um, you can see it's summarizing a video from How It's Made. That was one of my favorite videos as a kid growing up. So, getting started with this one is as simple as heading to the URL that I'm going to link around this video. And you can see we have an overview page here. This is a template, so all I have to do is click Use Template, and then, you know, give it a descriptive name, something different than the default maybe. Um, and that's going to fork the Replit into our own environment. So, it's going to create a copy under our Replit account. We own all the code here. Uh, we could link this to GitHub if we wanted to. We could download it as a zip, um, or take it off the platform if that's something you are.

interested in, but this is the uh repet workspace, and it might look familiar if you've tried uh other coding tools before if you're a developer, right um. The pain view is pretty uh reminiscent of something like a VSS code. The difference here though is that this lives in the cloud, and we have our own is ated environment, which means that when we make changes in this uh code editor, it's like having a folder with a virtual environment um or its own uh set of dependencies um managed for us uh by repet, but under the hood, it's all poetry, it's all python stuff since this template is in Python, which makes it completely compatible with any other tool that you might use. So, all we have to do here to get started is paste in a Gro API key, and I already have one, so I'm going to do that now uh. But if you haven't used Gro before, you can just head over to the grock website, that's gq.com uh, to configure your own, and it's a completely free service. There are some

rate limits, but Grock uses the llama 3 model, which is an open-source model, and that's how the service is free. They're a really great team. I love some of their demos, and they provide a lot of really great resources for getting started. But pasting that secret in, all we really have to do is click run, and that's going to run what we have configured. We're not installing any dependencies since the template already has dependencies installed, and in the web view, we're going to get the demo I walked through earlier. So yeah, it's that easy to get started with the summarizer. But now I'd like to take some time and actually walk through the code to help you understand what we're doing. So I expanded the main pane here to walk through the code, and I just really want to give a high-level overview of how this works if you'd like to make some changes or play around with this. So again, I use the Gradio framework, and this maybe these 20

lines of code here are all that instantiate a Gradio instance. So, one of the ways you can define a Gradio interface is using this with blocks syntax, and then I'm just defining some markdown, and then an interface. The interface is what contains our application. So, in the interface, a Gradio interface, you have inputs that go through a function and return outputs. Pretty logical, but our inputs are a text box with a URL. Our outputs are the markdown that you saw in our demo, and then our function is this Python function called app. And so, I use the app function really as a wrapper that all it does is take in our URL and then pass it to sort of a more main function I called summarize YT, and it's going to return each value using yield, which is the way you return a generator component in Python, unless there's a rate exception. And I use this main function basically for that rate limit functionality. We're getting our YouTube transcript and building metadata. Those are two

functions defined a little bit further up. You can take a look at those, but they just use a few open-source libraries I found for extracting that data given a URL. Um, but the important part is that we're defining a series of messages, and uh, the cool thing about uh, Gro, the cool thing about llama, is that the chat completion functionality is very similar to other AI libraries. So, if you want to do this in Gemini, if you want to do this in OpenAI or Anthropic, the syntax is very similar, and swapping out the models might take you 10 or 15 minutes. Um, but all I'm doing is chunking that transcript up. Uh, I know what the transcript, it's a list of dictionaries. I know what the transcript looks like. Um, so I'm doing a little bit of list and numpy Jiu-Jitsu uh, to get that chunked up there, and then for each chunk, we're going to summarize that transcript using our system prompt, which is defined at the top of the file. Um, and

this is built in such a way that if you wanted to change the number of chunks, you could. If you wanted to change CH, the system prompt, you could. So, I'd really implore you to tinker with this, see if you can make it better. I'm already thinking of ways to make this better. For example, a fixed number of chunks isn't a very good way to summarize a YouTube video. But what if we pass the transcript through a model before we summarized it to find the chapters for us? Or even to find a tool or a function that could do that in a structured way? Well, then we'd have something that could segment the video on the fly, delineate those transcripts, chunk them up by itself using AI, and do that work for us. But yeah, you know, that's at a high level how the app works. It takes that transcript, uses some Python functions to fetch it, chunks it up, and then summarizes it by those chunks. But I

think that there's a lot of room for improvement, so I'd love to see how you guys can improve on this application. Again, this is hosted on Replit. You might find that's a really easy way to build and share your apps, and if you do, I would love for you to share them with me. But I'm Matt. I work in developer relations at Replit, and this is my YouTube summarizer app. So, I'd love if you could give it a shot. Peace.

Deploy an AI YouTube Summary app with Groq and Llama 3

Scripts