RAG - JavaScript

JavaScript port of the RAG (Retrieval-Augmented Generation) assignment using Ollama.

Prerequisites

• Node.js 18+
• Ollama installed and running locally
• Required Ollama models pulled:
  • embeddinggemma (embeddings)
  • phi3:mini (chat generation)

Install

From this folder (Chapter-06-RAG/assignment/js):

npm install

Setup

1. Start Ollama if it is not already running.
2. Pull the required model:

ollama pull embeddinggemma

3. (Optional) Choose which markdown file to ingest by setting RAG_SOURCE_DOC.

PowerShell:

$env:RAG_SOURCE_DOC = "faq.md"

If not set, the default is faq.md.

4. (Optional) Choose chat model for answer generation:

$env:RAG_CHAT_MODEL = "phi3:mini"

5. (Optional) Force a fresh ingest by deleting the vector store file:

# PowerShell
Remove-Item ..\vector_store_faq.json -ErrorAction SilentlyContinue
Remove-Item ..\vector_store_products.json -ErrorAction SilentlyContinue

Run

npm start

Open http://localhost:3000.

The page includes:

• Products menu backed by a local SQLite catalog that is created automatically if missing
• FAQ page rendered from markdown (faq.md by default)
• Smart Search with Auto, Products, and FAQ modes
• Product search results that return a matched product and an add-to-cart link

On first run, catalog.js creates products.sqlite with five sample products if the database is missing. rag.js reads ../docs/<RAG_SOURCE_DOC> for FAQ content, reads products from SQLite, generates embeddings via Ollama, and saves them to ../vector_store_faq.json and ../vector_store_products.json. Subsequent runs load the stores from disk and skip ingestion unless the FAQ file or SQLite database changes.

Usage as a module

rag.js exports three functions you can import in other scripts:

import { ingest, getChunks, search } from "./rag.js";

// Ingest (or load) the FAQ and product vector stores
const stores = await ingest();

// Get all text chunks from the source document
const chunks = getChunks();

// Search for the most relevant FAQ or product match
const result = await search("your question here", stores, "auto");

It also exports answerWithRag(prompt, mode, stores) which returns an LLM-generated answer grounded in the retrieved FAQ chunk or product record.