Andrew Ng introduces a new course on function calling and data extraction with LLMs, taught by Jiantao Jiao and Venkat Srinivasan from Nexusflow. You'll learn how LLMs invoke external tools through structured outputs, build agent workflows with parallel and nested calls, and work with the open-source Nexus model to extract customer-service data into databases.
I'm delighted to introduce function calling and data extraction with LLMs built in partnership with Nexus Flow, taught by Jiantao Jiao, who is co founder and CEO of Nexus Flow, as well as Assistant professor of ECS and Statistics at UC Berkeley, and Venkat Srinivasan, who is founding engineer at Nexisflow Function calling is a core capability of AI agents as well as advanced chatbots. You may have noticed when, say, using ChatGPT that it might decide to call a service like Websearch to gather information to answer your query. This is called function calling. This course will teach you how to use function calling in your own applications. Here's how it works. The LLM has in its prompt a description of, say, a web search function, as was the format of the arguments of this function. When DLM recognizes a query that is best served by making that function call, then the LLM returns or generates a string that can be used to call that function, including the appropriate arguments like what terms to search on. Note that the LLM just forms a string it doesn't perform the function call itself. These functions are often referred to also as tools, and they're a powerful way to extend the capabilities of a chatbot and are a core component of AI agents, like maybe a research agent whose tools might include web search and Wikipedia lookups. Function calling requires taking a natural language input and outputting structured data. This capability turns out to be useful for many applications beyond chatbots and agents.
That's right, Andrew. There are many instances where you want to convert natural language inputs into structured outputs. For example, you may have a library of functions for business intelligence dashboard operations, and you wish to add a natural language interface to invoke them. Alternatively, you might have applications that process transcripts, notes, or proceedings and want to store them to a database. You do not need large general purpose LLMs to accomplish such tasks. For example, in this course you use the NextDriven v2.13B, an open source model available on HuggingFace. The model outperforms GPT4 in some function coding tasks and has only 13 billion parameters. A model this size can be hosted locally and used for your favorite applications. Venkat will show you how to do this.
In this course you will get lots of hands on experience with function calling. You will start with the basics with simple function calls and then move on to more complex situations like multiple parallel and nested function calls. These capabilities let you build much more complex agent workflows in which an LLM can plan and execute a sequence of function calls to accomplish a goal. You'll also learn to use open API specifications to build function calls that can access web services. And you'll wrap up by building an application which takes in customer service transcripts, processes them, and stores results in a database with commands generated by the LLM. Once you finish the course, you will be confident in using function calling in your LLM and agent applications.
That sounds great. Humans are much more powerful when we don't have only our bare hands, but are allowed to use tools. OMs and AI agents too are much more powerful when you give them access to tools. I hope you enjoy learning about these function calling and tool use design patterns and apply them to your own applications.
