What is LangChain?

What is LangChain?

Stop me if you've heard this one before, but there are a lot of large language models available today and they have their own capabilities and specialities. What if I prefer to use one LLM to interpret some user queries in my business application, but a whole other LLM to author a response to those queries?

Well, that scenario is exactly what LangChain caters to. LangChain is an open-source orchestration framework for the development of applications that use large language models, and it comes in both Python and JavaScript libraries. It's essentially a generic interface for nearly any LLM, so you have a centralized development environment to build your large language model applications and then integrate them with stuff like data sources and software workflows.

Now, when it was launched by Harrison Chase in October 2022, LangChain enjoyed a meteoric rise, and by June of the following year, it was the single fastest-growing open-source project on GitHub. And while the LangChain hype train has slightly cooled a little bit, there's plenty of utility here. So let's take a look at its components.

So what makes up LangChain?

The Core Concept: Abstractions

Well, LangChain streamlines the programming of LLM applications through something called abstractions. Now, what do I mean by that? Well, your thermostat that allows you to control the temperature in your home without needing to understand all the complex circuitry that this entails—we just set the temperature. That's an abstraction.

So LangChain's abstractions represent common steps and concepts necessary to work with language models, and they can be chained together to create applications, minimizing the amount of code required to execute complex NLP tasks.

The LLM and Prompt Modules

So let's start with the LLM module. Now, nearly any LLM can be used in LangChain; you just need an API key. The LLM class is designed to provide a standard interface for all models, so pick an LLM of your choice, be that a closed-source one like GPT-4 or an open-source one like Llama 2—or this being LangChain, pick both.

Okay, what else we got? We have prompts. Now, prompts are the instructions given to a large language model, and the PromptTemplate class in LangChain formalizes the composition of prompts without the need to manually hardcode context and queries. A prompt template can contain instructions like, "Do not use technical terms in your response," that would be a good one, or it could be a set of examples to guide its responses—that's called few-shot prompting—or it could specify an output format.

Composing Workflows with Chains

Now, chains, as the name implies, are the core of LangChain workflows. They combine LLMs with other components, creating applications by executing a sequence of functions. So let's say our application needs to first of all retrieve data from a website, then it needs to summarize the text it gets back, and then finally it needs to use that summary to answer user-submitted questions. That's a sequential chain where the output of one function acts as the input to the next, and each function in the chain could use different prompts, different parameters, and even different models.

Connecting to External Data with Indexes

Now, to achieve certain tasks, LLMs might need to access specific external data sources that are not included in the training data set of the LLM itself. So things like internal documents or emails, that sort of thing. Now, LangChain collectively refers to this sort of documentation as indexes, and there are a number of them. So let's take a look at a few.

One of them is called a document loader. Now, document loaders, they work with third-party applications for importing data from sources like file storage services—so think Dropbox or Google Drive—or web content from like YouTube transcripts, or collaboration tools like Airtable, or databases like Pandas and MongoDB. There's also support for vector databases as well.

Now, unlike traditional structured databases, vector databases represent data points by converting them into something called vector embeddings, which are numerical representations in the form of vectors with a fixed number of dimensions, and you can store a lot of information in this format as it's a very efficient means of retrieval.

There are also something called text splitters, which can be very useful as well because they can split text up into small, semantically meaningful chunks that can then be combined using the methods and parameters of your choosing.

Adding Memory and Reasoning with Agents

Now, LLMs by default don't really have any long-term memory of prior conversations unless you happen to pass the chat history in as an input to your query. But LangChain solves this problem with simple utilities for adding memory into your application. And you have options for retaining the entire conversation through to options to just retain a summarization of the conversation that we've had so far.

And then finally, the last one we'll look at are agents. Now, agents can use a given language model as a reasoning engine to determine which actions to take and when. When building a chain for an agent, you'll want to include inputs like a list of the available tools that it should use, the user input like the prompts and the queries, and then any other relevant previously executed steps.

Practical LangChain Use Cases

So how can we put all of this to work for our applications? Well, let's talk about a few LangChain use cases. Now, obviously, we have chatbots. LangChain can be used to provide proper context for the specific use of a chatbot and to integrate chatbots into existing communication channels and workflows with their own APIs.

We also have summarization. A language model can be tasked with summarizing many types of text, from breaking down complex academic papers and transcripts to providing just a digest of incoming emails. We've also seen lots of examples where this is used for question answering. So using specific documents or specialized knowledge bases, LLMs can retrieve the relevant information from the storage and then articulate helpful answers using the information that would otherwise not have been in their training data set.

And yeah, this is a good one: data augmentation. LLMs can be used to generate synthetic data for use in machine learning. So for example, an LLM can be trained to generate additional samples that closely resemble the real data points in a training data set.

And there are, of course, virtual agents, as we already started to discuss. Integrated with the right workflows, LangChain's agent modules can use an LLM to autonomously determine the next steps and then take the action that it needs to complete that step using something called RPA or robotic process automation.

The LangChain Ecosystem & Conclusion

LangChain is open source and free to use. There are also related frameworks like LangServe for creating chains as REST APIs and LangSmith, which provides tools to monitor, evaluate, and debug applications. Essentially, LangChain's tools and APIs simplify the process of building applications that make use of large language models.

If you have any questions, please drop us a line below. And if you want to see more videos like this in the future, please like and subscribe. Thanks for watching.