Build a RAG App with LangChain, OpenAI & Node.js — Step by Step

At India Today Group, we needed an intelligent search system that could understand editorial queries like "find articles about economic policy impact on rural India" — not just keyword matching, but semantic understanding. Here's how I built it using RAG (Retrieval-Augmented Generation).

What is RAG?

RAG combines a retrieval system (finding relevant documents) with a generation model (LLM like GPT-4) to produce accurate, context-aware answers grounded in your actual data — not hallucinated facts.

Architecture Overview

User Query → Embed Query (OpenAI) → Vector Search (MongoDB Atlas)
  → Top K Documents → LLM Prompt + Context → Generated Answer

Step 1: Setup Dependencies

npm install langchain @langchain/openai @langchain/community mongodb

Step 2: Document Ingestion Pipeline

First, we ingest articles from our CMS, split them into chunks, generate embeddings, and store in MongoDB Atlas Vector Search:

import { OpenAIEmbeddings } from '@langchain/openai';
import { RecursiveCharacterTextSplitter } from 'langchain/text_splitter';
import { MongoDBAtlasVectorSearch } from '@langchain/community/vectorstores/mongodb_atlas';

const embeddings = new OpenAIEmbeddings({
  openAIApiKey: process.env.OPENAI_API_KEY,
  modelName: 'text-embedding-3-small',
});

const splitter = new RecursiveCharacterTextSplitter({
  chunkSize: 1000,
  chunkOverlap: 200,
});

// Split articles into chunks
const docs = await splitter.createDocuments(
  articles.map(a => a.content),
  articles.map(a => ({ title: a.title, id: a._id, date: a.publishedAt }))
);

// Store with embeddings in MongoDB
await MongoDBAtlasVectorSearch.fromDocuments(docs, embeddings, {
  collection: mongoCollection,
  indexName: 'article_vector_index',
});

Step 3: Query Pipeline with LangChain

import { ChatOpenAI } from '@langchain/openai';
import { RetrievalQAChain } from 'langchain/chains';

const llm = new ChatOpenAI({
  modelName: 'gpt-4-turbo-preview',
  temperature: 0.2,
});

const vectorStore = new MongoDBAtlasVectorSearch(embeddings, {
  collection: mongoCollection,
  indexName: 'article_vector_index',
});

const chain = RetrievalQAChain.fromLLM(llm, vectorStore.asRetriever(5));

// Query
const result = await chain.call({
  query: 'What are the latest developments in India education policy?',
});
console.log(result.text); // AI-generated answer with citations

Step 4: Production Considerations

In production at India Today, we added: caching (Redis for repeated queries), rate limiting, streaming responses for real-time UX, and source attribution so editors can verify AI-generated summaries.

Results

The RAG system reduced editorial research time by 60% and powers the intelligent search across India Today's digital platform. Editors can now ask natural language questions and get accurate answers grounded in our 50,000+ article archive.