Sign in to continue learning
Jupyter AI: AI Coding in Notebooks
Beginner
44m
Topics
AI Coding
In this lesson, we'll look at more advanced ways to use Jupyter AI as a coding assistant. The motivating example is building a book recommendation assistant using the Open Library API, which returns information about books. For APIs that LLMs aren't already trained to use, one key idea is writing documentation for that API, which you'll see me do in a markdown file. So you can then give that markdown file as context to the LLM, so that it gets the API usage instructions it needs to generate correct code. You also see some of the best practices for organizing context for AI coding. Let's take a look. In this lesson, we're going to use the Open Library API in order to build an AI book research assistant. And in doing so, we'll go over some best practices. for using Jupyter AI to code for you. We'll be using the Internet Archive's Open Library APIs to search for books. The Open Library has a set of different APIs, for example, to search for books with the Book Search API and a number of others. In order to make our LLM know how to use these APIs, I've created here a markdown doc. Let's open that up. Open with Markdown Preview. that is a file with detailed documentation on how to use these APIs or how to use these different endpoints. Feel free to read through this documentation, but the one I'm going to use is actually this one, which is the Search API. that lets you, for example, search for Lord of the Rings or Pride and Prejudice or books by Tolkien or search for fantasy books. and so on. Feel free to browse this markdown file yourself after you're done watching this video and read through some of this documentation if you wish. I'm going to click on the bubble here to start the chat and um, let's have a Book assistant chat. And what I'm going to do is attach my openlibrary_api_docs.md markdown file, I would then type @file or you can actually click here and then um... I'm going to type Open Library, there pops up and click this. So I'm attaching this markdown file's documentation. Then here's the problem we're going to use. Write a code cell that uses Open Library's search endpoint to look up books and so on. Uh returns a response list, listing the key, title, author_name, and so on. And then in this case, let's limit it to five results. And I'm saying, use the Open Library API documentation in the attached file. So let's run that. and it generates this piece of code. So let me copy it over to my notebook with one click and um, let's give this a shot. Enter your book search query. I want to find science fiction. And so this then I guess calls the API and gets back the top five results. So Science Fiction Hall of Fame, Robert Silverberg, Heinlein, Arthur C. Clarke, Isaac Asimov, lots of great authors. Oh, the Foundation novel, which I love, by Isaac Asimov. Um, Stranger in a Strange Land and you know. bunch of uh pretty interesting science fiction books. So it looks like this is working. In order to build a book research assistant, what I'd like to do is take this code, this Jupyter Notebook cell, it just calls the API and prints out the results. Take this code and wrap it in a function so that I can then give this as a function or as a tool to an LLM to call. One way I could do that is I could drag this cell. You know, so I'm clicking out here, drag it over, and uh attach this and then say, please take this code cell and wrap it in the function. Another way without attaching it is write a prompt that says, wrap the code above in a function search_books with good docstring documentation. inputs are the query and the limit. Let's now have this default to returning 10 books and output a JSON string and so on. And let's run that. And then now, um, takes the code and looks like it did a decent job with, wow, pretty nice documentation in fact. So let me now take this code and um, copy paste Jupyter Notebook. So I can just define it as follows. And if I want to test it, I can then try results = search_books("science fiction"), print(results). And now it's returning the top 10 books. Okay, so that works. So we've successfully created a search_books function that uses the Open Library API to search for books. Next, we need to give this function as a tool to an LLM to build the book research assistant. In order to do that, we're going to use the open source aisuite library, which is a library that makes it easy to get um to use tools. as well as makes it easy for you to switch between different LLM providers. This is an open source package that a friend Rohit Prasad and I have been working on. But because it's a fairly new package, LLMs have not yet been natively trained on the syntax of how to call aisuite. So one thing we could do is write a AI aisuite API documentation markdown file like we did with the Open Library API docs markdown file. But in order to use aisuite successfully, it turns out giving it a code example like this should be enough to get the LLM to figure it out. So I've written here in a raw Jupyter markdown cell, just an example of how to use aisuite. And uh if you're familiar with tool calling or function calling using an LLM, aisuite has a relatively simple library where you just have to tell it what are the tools, which is search_books in this example, and um, this code snippet is enough to get a LLM to decide if it needs to call the search_books function, which we just created up above in order to fulfill a user query. So what I'm going to do is let's take this cell, drag it to the chat, and then I'm going to prompt the LLM. Use this example to create a simple chat handler. function chat_with_tools, uh input the user_message and then output that. And it can use the search_books tool if it wishes. So this generates the code, and it looks like it's giving the search_books function we created above as a tool to the AI correctly. So let's copy this over. And um, let's go ahead and run it. So that has defined the chat_with_tools function. And now here's a Jupyter notebook cell where my query is search for books about science fiction. Let's run that. And so we're going to use the chat_with_tools function that we defined above. So this runs for a minute and it looks like it calls a tool and gets a decent set of results. And these are the results returned by the function calling. So these results look decent, but for debugging purposes, I want to be sure that this really used the function we defined up above rather than just use the LLM's pre-trained knowledge to regurgitate a bunch of plausible sounding books. So let me go back up to this function, and I want to add a print statement to just let me know a status update when this function is being used. I'm going to tell it modify the search_books function to add a print statement. showing when it is called with basic status messages. All right. So that should make the operation of this a little bit more transparent. It's generating the code and um I'm going to select the cell and let's just replace it. It's going to redefine this. Um, let's see what happens this time around. Okay, good. Now it's uh giving me a bunch of status updates before it... has returned. And then if I rerun this, time let's look for books on dragons. Search for books about dragons. Oh, I like dragons. Let's see what happens. Good. So it's looking for books on dragons. Right, some of these look like they are pretty fun. So that shows how you might take a cell and ask the chatbot to iterate and modify the code for you as well. After this video, Please go to the Jupyter Lab on this website and play with this yourself. And in that lab, we have some optional instructions you can follow if you want to use some of these other APIs to build potentially more complex book research agents. such as maybe after retrieving certain books, you might want to look up some details in that work to print out more details about the books that was retrieved. So, please go to the lab and I hope you have fun using Jupyter AI to build applications like this. And uh, maybe you find some cool books you want to read too.
