Build a RAG pipeline with LlamaIndex

Prerequisites

Step-by-Step

1. 1

   Install LlamaIndex.TS

   The TypeScript SDK ships ESM-first. Make sure your tsconfig has module: nodenext.

   pnpm add llamaindex

2. 2

   Load documents

   SimpleDirectoryReader auto-detects file types. For PDFs, install pdf-parse alongside.

   import { SimpleDirectoryReader, VectorStoreIndex } from 'llamaindex';
   
   const reader = new SimpleDirectoryReader();
   const documents = await reader.loadData({ directoryPath: './docs' });

3. 3

   Build the index

   VectorStoreIndex.fromDocuments handles chunking, embedding, and vector storage in one call. Persist it to disk so you do not pay for re-embeds.

   import { storageContextFromDefaults } from 'llamaindex';
   
   const storageContext = await storageContextFromDefaults({ persistDir: './storage' });
   const index = await VectorStoreIndex.fromDocuments(documents, { storageContext });

4. 4

   Create a query engine

   The query engine wires retriever, response synthesizer, and post-processors together. Use similarityTopK to control recall.

   const queryEngine = index.asQueryEngine({ similarityTopK: 5 });
   const response = await queryEngine.query({ query: 'Summarize the onboarding flow' });
   console.log(response.toString());

5. 5

   Add metadata filters

   Tag documents at ingest time so you can scope queries to a subset. This is where LlamaIndex shines vs raw vector stores.

   documents.forEach((d) => { d.metadata = { ...d.metadata, tenant: 'acme' }; });

6. 6

   Stream responses

   Streaming makes the app feel instant. Pipe tokens to your UI as they arrive.

   const stream = await queryEngine.query({ query: 'Explain pricing', stream: true });
   for await (const chunk of stream) process.stdout.write(chunk.response);

Common Pitfalls

What's Next

• ->Switch storage to Postgres + pgvector for production.
• ->Add a routing query engine to fan queries across multiple indices.
• ->Wrap the engine in a chat agent for multi-turn use.

More Build a RAG pipeline