Vertex AI RAG with Node

This project implements a Retrieval-Augmented Generation (RAG) system using Google's Vertex AI and Gemini models.

How RAG Works

1. Document Processing: The PDF is chunked into smaller pieces (1000 characters each)
2. Embedding Generation: Each chunk is converted to a vector embedding using text-embedding-004
3. Vector Storage: Embeddings are stored in embeddings.json with their corresponding text
4. Query Processing: When you ask a question:
   • The question is converted to an embedding
   • Cosine similarity search finds the most relevant chunks
   • Top 3 most similar chunks are retrieved
   • Context is passed to Gemini 2.0 Flash to generate an answer

Setup

1. Install dependencies:

   npm install

2. Set up Google Cloud authentication:

   gcloud auth login

3. Create .env file:

   GOOGLE_CLOUD_PROJECT=YOUR_GCP_PROJECT_ID
   GOOGLE_CLOUD_LOCATION=global

Usage

1. Run the Demo

The simplest way to use the system is to run the demo:

node index.js

This will:

• Check if embeddings.json exists
• If not, process grad.pdf and create embeddings
• Ask a sample question about Stepwik founders
• Display the answer

2. Manual Usage

You can also use the functions directly:

// Create embeddings (run once)
await createEmbeddings();

// Ask questions
const answer = await getAnswer("What is Stepwik?");
console.log(answer);

Files

• index.js - Main RAG implementation with two functions:
  • createEmbeddings() - Processes PDF and creates embeddings
  • getAnswer(question) - Queries the RAG system
• embeddings.json - Stores embeddings and text chunks
• grad.pdf - Your source document

Configuration

The system is configured for:

• Project: agprop-in
• Location: global
• Embedding Model: text-embedding-004
• Generation Model: gemini-2.0-flash
• Chunk Size: 1000 characters
• Top Results: 3 most similar chunks

Example Output

🎯 RAG DEMO - Two Simple Functions
==================================================

✅ Embeddings already exist, skipping creation

2️⃣ Now asking questions...

----------------------------------------
🤔 Question: How are the founder of Stepwik?
📚 Loaded 15 embeddings
🔍 Most relevant chunks found:
  1. Similarity: 0.823
  2. Similarity: 0.712
  3. Similarity: 0.689
✅ Answer generated!
💡 Answer: Based on the context, the founders of Stepwik are...

Key Features

• Automatic Setup: Checks for existing embeddings and creates them if needed
• Cosine Similarity: Uses mathematical similarity scoring for accurate retrieval
• Rate Limiting: Includes delays to avoid API rate limits
• Error Handling: Graceful error handling with informative messages
• Context-Aware: Provides relevant document context to the AI model

This ensures the model answers based on your specific document content rather than general knowledge.