ChatGPT caught global attention at the end of 2022. Two years later, this AI-powered LLM gets 133 million visitors every month. So it's no wonder why developers are racing to incorporate this transformative technology into their products.
LangChain.js is a simple, modular framework with tools for developing applications powered by large language models. Build conversational AI, automate workflows, plug in context-aware retrieval strategies, and generate content that supports your application's cognitive architecture.
LangChain's open-source framework supports building production-ready AI apps with multiple value-added features for JavaScript-based applications. This reduces the time required to bring your AI apps from vision to prototype to production.
In this article, we’ll look at what LangChain is, the benefits of using LangChain JS, and how you can use it today to create high-performance GenAI apps that weave together large language models (LLMs) with external data sources.
Understanding LangChain and LangChain JS
Leveraging an LLM, you can quickly build GenAI applications based on existing foundational models. Combining calls to LLMs with your own data using techniques such as retrieval-augmented generation (RAG) provides the LLM with recent, domain-specific data, resulting in more relevant and accurate output for your users.
However, building reliable AI applications requires more than making a REST API call to a LLM. For example, you may need to interface with different LLMs based on their performance and how accurately they respond to specific queries. Then there’s all the supporting infrastructure your app might need, such as error handling, parallelization, building agents, and logging/monitoring.
LangChain is a framework, available for both Python and JavaScript, for developing applications that incorporate LLMs. The framework consists of various open-source libraries and packages that work together to enhance the application lifecycle for language model-driven tasks. Using LangChain, you can code GenAI apps that leverage LLMs, vector databases, and other critical components without coding out all the little details yourself.
LangChain JS brings the power of the LangChain framework to JavaScript applications. LangChain JS focuses on simplifying GenAI app development in three key areas:
Development. LangChain JS supports a number of out-of-the-box components for assembling GenAI apps quickly. You can add features such as streaming, route execution, message history, text splitting, parsing structured output, and more with less code than you’d use implementing everything on your own. You can also use LangGraph.js to build out stateful agents.
Productionization. LangChain JS supports LangSmith, which provides monitoring capabilities you can use to assess and fine-tune your app’s performance. LangSmith also supports an API you can call from the non-LangChain parts of your application.
Deployment. You can use LangGraph Cloud to turn your LangChain JS app into a callable, production-ready API that’s horizontally scalable and uses durable storage.
Components of LangChain and JS
Before we dive into getting started with LangChain JS, let’s look at the components that LangChain supports. These components serve as building blocks for constructing your app using a simple declarative language.
Here are a few of the key components in LangChain JS you should understand before using it. For other components, see the LangChain JS conceptual guide.
LLMs and chat models
Most modern LLMs operate as chat completion models - i.e., they take a series of messages as inputs and return generated chat messages as output. Responses are returned as messages that describe the role the chat model is assuming as well as the context of the message.
LangChain JS supports dozens of chat models and LLMs out-of-the-box, making it easy to test different models or leverage different models for different purposes. For example, you may use one model for user-generated chat and another model for creating structured JSON output.
Agents
One of LangChain’s more powerful features, agents provide decision-making capabilities, using the output from an LLM to determine the next action in the chain. You can use LangGraph.js to create complex agent and stateful multi-actor applications with support for cycles, controllability, and persistence.
Prompt templates
Prompt templates take user or input and supply them as part of the instructions you send to a language model. Prompt templates support generating multiple message types and prompt outputs.
Chains
LangChain JS supports the LangChain Expression Language (LCEL), which defines a simple, declarative way to chain LangChain components together into a full, working AI application.
Using LCEL, you can quickly create a prototype app that includes anywhere from a few steps to over 100. You can then ship that prototype to production with few code changes. For example, the following call pipes a parameterized prompt template to an LLM and then pipes the output to a parser that reads and formats the response:
const chain = promptTemplate.pipe(model).pipe(parser);
LCEL and chains provide much of the value of LangChain JS. Using LCEL gives you streaming support, parallelization, retry and fallback logic, access to intermediate results, and tracing via LangSmith. That’s functionality you don’t have to develop and test yourself.
Tool calling and output parsers
You can use tool calling to respond to a given prompt by generating output in conformance with a user-defined schema. LangChain JS support tool calling for all LLM providers that support it, providing a consistent interface across the models.
For models that don’t support tool calling, LangChain JS provides parsing via output parsers. LangChain supports built-in parsers for all popular structured text formats.
Setting up LangChain JS
To get started with LangChain JS, install Node.js version 18 or higher, as well as one of the supported package managers - npm, Yarn, or pnpm. Then, install the LangChain JS library:
npm i langchain @langchain/core
(For non-npm examples, see the LangChain JS introduction on their website.)
Next, install the integrations you want to use. This will be the LLMs you’ll leverage in your app. For example, to install support for OpenAI, run:
npm install @langchain/openai
You can also install both packages simultaneously:
npm install @langchain/openai @langchain/core
Finally, perform any additional setup that your integrations require. The common example is setting an API key for your desired LLM. For example, for OpenAI, after creating a project and adding an API key, you can use it by setting the OPENAI_API_KEY variable:
OPENAI_API_KEY=your-api-key
(Note: It’s generally recommended to set this via code after retrieving it from a secrets vault supported by your cloud provider, such as AWS Secrets Manager, GCP Secret Manager, or Azure Key Vault.)
Using LangChain JS
What you do next with LangChain JS depends on what you want to build. Here are a few examples of what you can create with LangChain JS and the capabilities you can leverage.
Create a custom chatbot
To create a basic chatbot, you can instantiate a model via LangChain using the following code:
import { ChatOpenAI } from "@langchain/openai";
const model = new ChatOpenAI({
model: "gpt-4o-mini",
temperature: 0
});
At its most basic, you can create a chatbot by making a direct call to the LLM:
await model.invoke([new HumanMessage({ content: "Hi! I'm Bob" })]);
You can then build on this to add more features and knowledge context. For example, you can leverage the message history feature in LangChain JS to save chat history to a data store, reload it in future sessions, and provide it to the LLM as context. That way, when the user next engages with the chatbot, the LLM can incorporate knowledge from past conservations into its responses.
Customizing your chatbot with RAG
LLMs provide great foundational language comprehension and generation abilities, coupled with a general knowledge based on large data sets. Their limitation is that they’re a snapshot in time. Their data is limited to the last time they were trained (usually a year or more old).
Additionally, your company has volumes of proprietary, domain-specific data - e.g., past customer support transcripts, product model information and documentation, etc. Sending this information to an LLM as part of the prompt ensures more relevant, accurate, and up-to-date output.
Retrieval-augmented generation, or RAG, incorporates external data in real-time based on user input. It incorporates information from documents, API calls, and other external data sources and feeds them to the LLM as part of the prompt.
One common implementation of RAG is to store information in a vector database, which enables fast, complex, and context-sensitive searches across diverse data sets. LangChain JS works integrated with multiple vector database systems so you can include RAG-derived data in your prompts with a few lines of code.
LangChain JS works with our own vector database solution, Astra DB, which provides data search with petabyte scalability. We provide a detailed code example that shows you how to import documents into Astra DB, search them on-demand with a user’s query, and then embed the results as context into your LLM call.
Other LangChain JS capabilities
You can build more than chatbots with LangChain JS. You can use it to create sentiment analysis engines, structured content generation pipelines, and more.
LangChain JS also supports several third-party integrations besides LLMs - e.g., Redis for in-memory vector storage. It also supports multiple tools you can use to pull data from search engines, interpret code, automate tasks, and integrate with other third-party data sources such as Twilio, Wolfram Alpha, and more.
Best practices for using LangChain JS
Using a framework like LangChain JS doesn’t eliminate the need to fine-tune your AI app for your specific use case. A few best practices include:
- Use LangSmith to trace and benchmark request/response times for your app.
- Use langchain-benchmarks or a similar benchmarking tool to assess your LLM’s performance for specific task types. LangChain Benchmarks ships with four environments you can use to stress-test your agents.
- Consider leveraging a human-in-the-loop approach for your AI use case. Trying to automate everything can result in decreased accuracy or even outright hallucinations. Adding a human expert into the equation provides the best of both worlds, using AI technology to augment human intelligence rather than replace it.
- Prioritize security! The basics of good security apply to an AI app as they would with any app in your software portfolio. Use the principle of least privilege when granting permissions to third-party systems, develop and run security tests frequently, and take a defense-in-depth approach to avoid issues such as data leaks and data breaches.
- Create a comprehensive compliance strategy to avoid issues such as leaking personally identifiable information (PII) or other highly sensitive information from your AI-powered applications.
Use LangChain JS and Astra DB
You can leverage the numerous tools and capabilities that LangChain JS exposes to build your GenAI application quickly from reusable building blocks. With our Astra DB LangChain JavaScript integration, you can easily store and query documents to increase the accuracy, timeliness, and relevance of LLM responses.
Get started today - sign up for a free Astra DB account and build your first LangChain JS app!
