Learning LangChain

Chapter 5. Cognitive Architectures with LangGraph

So far, we’ve looked at the most common features of LLM applications:

• Prompting techniques in the Preface and Chapter 1

• RAG in Chapters 2 and 3

• Memory in Chapter 4

The next question should be: How do we assemble these pieces into a coherent application that achieves the goal we set out to solve? To draw a parallel with the world of bricks and mortar, a swimming pool and a one-story house are built of the same materials, but obviously serve very different purposes. What makes them uniquely suited to their different purposes is the plan for how those materials are combined—that is, their architecture. The same is true when building LLM applications. The most important decisions you have to make are how to assemble the different components you have at your disposal (such as RAG, prompting techniques, memory) into something that achieves your purpose.

Before we look at specific architectures, let’s walk through an example. Any LLM application you might build will start from a purpose: what the app is designed to do. Let’s say you want to build an email assistant—an LLM application that reads your emails before you do and aims to reduce the amount of emails you need to look at. The application might do this by archiving a few uninteresting ones, directly replying to some, and marking others as deserving of your attention later.

You probably also would want the app to be bound by some constraints in its action. Listing those ...