Prompt Engineering for LLMs

Chapter 8. Conversational Agency

In Chapter 3, we covered the departure from text completion models to chat models. A chat model by itself is aware of only the information covered in training and whatever information the user has just told it. The chat model is unable to reach out into the world and learn about information that was unavailable during training, and it’s unable to interact with the world and take external actions on behalf of the user.

The LLM community is making great headway in overcoming these limitations through conversational agency. Agency is the ability of an entity to complete tasks and achieve goals in a self-directed and autonomous manner. The conversational agents that we discuss in this chapter provide an experience similar to chat—a back-and-forth dialogue between a user and an assistant—but add in the ability for the assistant to reach out to the real world, learn new information, and interact with real-world assets.

In this chapter, we’ll introduce several state-of-the-art approaches to building an LLM-based conversational agent. We’ll explore how models can use tools to reach out into the external world, how they can be conditioned to better reason through their problem space, and how we can gather the best context to facilitate long or complex interactions. By the end of this chapter, you’ll be able to build your own conversational agent that’s capable of going out into the world and performing guided tasks on your behalf.

Tool Usage

Working in ...