To be able to work with the code samples in this chapter, install the langchain-core library by running pip install langchain-core in your terminal.

Large language models (LLMs) are the fundamental components required to develop generative AI applications. However, as we will discuss in this book, these applications needed to be extended with state management, retrieval of supplementary data, interactions with the API, and so on. Such extended systems are often referred to as LLM-augmented autonomous agents (LAAs), which are computer programs capable of interacting with their environment and solving complex tasks with their underlying LLM by leveraging past observations and actions [1]. LAAs need orchestration, which is a process through which the LAA controls the underlying LLMs and coordinates the planning and step-by-step execution of a task.

LangChain is one of the most popular open source frameworks for LLM orchestration. Harrison Chase launched it as an open source project in 2022, and it attracted a lot of attention and quickly grew in popularity [2]. In February 2024, the LangChain team raised a Series A round...

As we have discussed, LangChain is a framework for orchestrating your generative AI applications. Let’s look at two key LangChain primitives: runnables and chains.

The most basic LangChain abstract interface is langchain_core/runnables/base.Runnable, and it makes it easy to compose different pieces together for the orchestration [4]. A Runnable defines three main methods (there are some more, mainly for asynchronous applications; you can take a look at the full documentation if you’re interested [4]):

• invoke calls Runnable on an input. It’s an abstract method that should be defined by the specific implementation.
• stream calls Runnable on an input and streams chunks of the response. With a default implementation, it just calls the invoke method and yields the output, but if you need proper streaming support, your Runnable should implement this method.
• batch calls Runnable on a list of inputs ...

Let’s have a high-level look at key LangChain building blocks. If you need a refresher on terms such as vectorstore, or embeddings are new to you, feel free to check the Index where you can find more details and links to relevant chapters with more detailed explanations.

Data structures

First, let’s explore some key data structures that you might encounter while developing applications. LangChain uses Pydantic, a Python data validation library; therefore, almost all LangChain data structures are Pydantic base models [7]. For those of you who are new to Pydantic, it’s a data validation Python library that allows you to define your data schemas, and if you haven’t heard about it, then you probably used an alternative schema definition library – dataclasses.

Note that LangChain has also its own langchain_core.load.Serializable class that inherits from Basemodel and adds additional control over serialization of...

To get started, you need to have a Google Cloud project with billing and enable the Vertex AI API on it. If you are new to Google Cloud or you’re unfamiliar with some of the Google Cloud services that we mention during this section, take a look at Chapter 4.

You must authenticate yourself on Google Cloud. Most probably, you’ll be trying things out with a Jupyter notebook (and we’ll look into deploying your application to production in Chapter 16). In that case, you have a few options:

• If you’re running a Jupyter notebook locally, you can use your local credentials. All you need to do is run the following in the terminal:

  gcloud auth application-default login

• If you’re running a Colab notebook, you can authenticate with your own credentials (but you’ll need to refresh them each time after your instance is stopped):

  from google.colab import auth
  auth.authenticate_user()

If you’re using a Jupyter...

What is an integration? LangChain exposes a set of interfaces, and different libraries or vendors can integrate with LangChain through their own implementation of these interfaces (of course, they might support additional parameters or features that other providers don’t have). The most obvious example of an interface is an interface for an LLM itself (we’ll discuss this in more detail in Chapter 3). You can think about your LangChain code as being a stateless transformation, and everything that requires I/O or state preservation (e.g., working with your data or storing the history of a chat conversation) will connect to different providers of such services (LLMs, databases, document storage, etc.) through an integration.

The open source community and Google continuously work together on bringing the LangChain experience to Google Cloud services:

• langchain-google-vertexai – Contains integrations that use Vertex AI (a...

In this chapter, we look into the landscape of LangChain products and libraries. We explained that LangChain is an open source framework for developing generative AI applications, and how they integrate with service providers. We also briefly looked at LangChain itself, the start-up behind this great framework.

Then, we discussed key LangChain primitives – the data structures that are based on Pydantic models and the main interfaces that are used in chains. We looked into what a chain is and at key primitives of LCEL – a declarative language to develop new chains. Now, you should be able to easily read LangChain code and understand what’s happening behind the scenes.

Finally, we looked at the integrations available with Google products, and we’ll explore their practical usage throughout the book.

Having the big picture in front of us, let’s start to dive deeper. In the next chapter, we’ll look into the basics of how to use...

1. BOLAA: Benchmarking and Orchestrating LLM-augmented Autonomous Agents, Z. Liu et al., 2023:

   https://arxiv.org/abs/2308.05960

2. LangChain, Wikipedia:

   https://en.wikipedia.org/wiki/LangChain

3. Announcing the General Availability of LangSmith and Our Series A Led By Sequoia Capital, Langchain blog:

   https://blog.langchain.dev/langsmith-ga/

4. LangChain Expression Language Interface, Langchain:

   https://python.langchain.com/docs/expression_language/interface

5. Evaluation order, Python Language Reference:

   https://docs.python.org/3/reference/expressions.html#evaluation-order

6. Service accounts overview, Google Cloud documentation:

   https://cloud.google.com/iam/docs/service-account-overview

7. Pydantic, product documentation:

   https://docs.pydantic.dev/latest/

8. Gemma: Introducing new state-of-the-art open models, J. Banks, T. Warkentin:

   https://blog.google/technology/developers/gemma-open-models/