In this introductory section, we outline the agenda and prerequisites for the LangChain course. We provide an overview of what LangChain represents and how it’s been used in real-life use cases.

The smallest building block of a language model is the token. The collection of all unique tokens represents the vocabulary of a language model. In this section, we’ll discuss how a token is defined in the context of OpenAI’s language models and how tokens determine a model’s price. Additionally, we’ll spend a few minutes presenting the different models OpenAI offers. Understanding token management is vital for optimizing the use of OpenAI models.

This section lays the groundwork for the course’s practical segment. We’ll cover creating a separate virtual environment with its own Python version and package set. Next, we’ll explain API key management and how to acquire an OpenAI API key—enabling our projects to use OpenAI’s expansive language models. Lastly, we’ll demonstrate setting this API key as an environment variable.

Before discussing the LangChain framework, we’ll first explore the OpenAI API basics—using various roles in prompts, setting the model temperature to adjust response randomness, limiting completion tokens, and streaming responses. This will later serve as a steppingstone towards understanding LangChain’s integration with OpenAI.

We now enter the realm of LangChain, starting with model inputs. We’ll distinguish between prompts, prompt templates, and chat prompt templates and explore the few-shot prompting—a technique that enhances model responses by training on tailored question-answer examples. We’ll also introduce the invoke() method—essential for the upcoming LangChain Expression Language section.

Language models are naturally stateless and don't retain previous interactions. This section of our how to build a chatbot tutorial examines popular methods for conversation storage—using a buffer, a capped buffer, or summarizing ongoing discussions and updating the summary after each interaction.

As we’ve seen in previous sections, the output from a language model in LangChain comes as an AI message. This format, however, often doesn’t come in handy as we can’t pass it to other methods or applications. Therefore, LangChain offers what is known as output parsers—making it possible to parse an AI message to a different object: a string, a CSV, a DateTime object, or whatnot.

We now arrive at the crux of our practice journey—a discussion allowing us to build the most essential objects in LangChain: the chains. The foundational idea behind the LangChain Expression Language is that the output of one chain component serves as the input to the next one, making the chains modular objects that can be quickly composed and combined. Additionally, we’ll discuss Runnable objects—the type of objects that can be entered as chain components.

In this course section, we address retrieval augmented generation—a technique that efficiently supplies context-specific information to a model. This process involves loading and segmenting personal data, embedding these segments, storing them in a vector store, and selectively retrieving and delivering only the relevant segments to the model. We’ll start with a theoretical discussion on this topic, then apply our knowledge to develop a basic 365 Q&A chatbot using transcripts from two 365 lectures.

The Build Chat Applications with OpenAI and LangChain Course's final section explores how reasoning chatbots select and apply the appropriate tools to tackle complex problems. This is accomplished by equipping a model with various tools, including a decision-making agent for tool selection and an executor to apply the chosen tool. Our LangChain course’s model will demonstrate this with a Wikipedia tool, a retriever tool, and a custom Python function.