Retrieval-Augmented Generation with LangChain.js

LangChain.js ships the same RAG building blocks as its Python counterpart: text splitters, embeddings, vector stores, and retrievers, all typed and composable with .pipe().

Splitting text

TypeScript

import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';

const splitter = new RecursiveCharacterTextSplitter({ chunkSize: 200, chunkOverlap: 20 });
const chunks = await splitter.splitText('LangChain is a framework for building applications with LLMs...');
console.log(chunks.length, chunks[0]);

import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';

const splitter = new RecursiveCharacterTextSplitter({ chunkSize: 200, chunkOverlap: 20 });
const chunks = await splitter.splitText('LangChain is a framework for building applications with LLMs...');
console.log(chunks.length, chunks[0]);

Like its Python counterpart, RecursiveCharacterTextSplitter prefers splitting on paragraph and sentence boundaries before falling back to a fixed size, keeping chunks semantically coherent, chunkOverlap repeats a few characters between adjacent chunks so a fact near a boundary isn't cut cleanly in half.

Embedding and storing

TypeScript

import { OpenAIEmbeddings } from '@langchain/openai';
import { MemoryVectorStore } from 'langchain/vectorstores/memory';

const embeddings = new OpenAIEmbeddings({ model: 'text-embedding-3-small' });
const vectorstore = await MemoryVectorStore.fromTexts(
  chunks,
  chunks.map((_, i) => ({ id: i })),
  embeddings
);
const retriever = vectorstore.asRetriever({ k: 3 });

import { OpenAIEmbeddings } from '@langchain/openai';
import { MemoryVectorStore } from 'langchain/vectorstores/memory';

const embeddings = new OpenAIEmbeddings({ model: 'text-embedding-3-small' });
const vectorstore = await MemoryVectorStore.fromTexts(
  chunks,
  chunks.map((_, i) => ({ id: i })),
  embeddings
);
const retriever = vectorstore.asRetriever({ k: 3 });

This needs a real API key to compute embeddings, so it's read-only here. MemoryVectorStore.fromTexts takes the chunks, per-chunk metadata (here, just an id), and the embeddings model, embedding every chunk and keeping the vectors in memory for similarity search. .asRetriever({ k: 3 }) wraps it as a Runnable returning the 3 most similar chunks for a given query.

Wiring retrieval into a chain

TypeScript

import { ChatPromptTemplate } from '@langchain/core/prompts';
import { StringOutputParser } from '@langchain/core/output_parsers';
import { RunnableSequence, RunnablePassthrough } from '@langchain/core/runnables';
import type { Document } from '@langchain/core/documents';

const prompt = ChatPromptTemplate.fromTemplate(
  'Answer the question using only this context:\n{context}\n\nQuestion: {question}'
);

const formatDocs = (docs: Document[]) => docs.map((d) => d.pageContent).join('\n\n');

const ragChain = RunnableSequence.from([
  { context: retriever.pipe(formatDocs), question: new RunnablePassthrough() },
  prompt,
  model,
  new StringOutputParser(),
]);

const answer = await ragChain.invoke('What is LangChain?');

import { ChatPromptTemplate } from '@langchain/core/prompts';
import { StringOutputParser } from '@langchain/core/output_parsers';
import { RunnableSequence, RunnablePassthrough } from '@langchain/core/runnables';
import type { Document } from '@langchain/core/documents';

const prompt = ChatPromptTemplate.fromTemplate(
  'Answer the question using only this context:\n{context}\n\nQuestion: {question}'
);

const formatDocs = (docs: Document[]) => docs.map((d) => d.pageContent).join('\n\n');

const ragChain = RunnableSequence.from([
  { context: retriever.pipe(formatDocs), question: new RunnablePassthrough() },
  prompt,
  model,
  new StringOutputParser(),
]);

const answer = await ragChain.invoke('What is LangChain?');

Read-only, needs a real model and vector store. RunnableSequence.from([...]) is the explicit, array-based way to build a multi-step chain, the object in the first step runs both branches on the same input in parallel: retriever.pipe(formatDocs) turns the question into formatted context, while new RunnablePassthrough() forwards the original question unchanged, exactly mirroring the Python version's dictionary syntax.

WARNING

"Using only this context" in the prompt is what keeps the model grounded in your retrieved documents instead of answering from its own training data, drop that instruction and you've built a chatbot that ignores the very documents you retrieved.

Retrieval-Augmented Generation with LangChain.js

Splitting text

Embedding and storing

Wiring retrieval into a chain

📝 LangChain.js RAG Quiz