The "Crawl, Walk, Run" of Implementing a RAG System

In this GitHub repo, I'm going to be using the same dataset and build a Human Resources RAG Agent three different ways: from the very basics of of RAG running on your local laptop using mainly open-source tooling, to fully managed platform on Google Cloud's Vertex AI. Welcome to my "Crawl, Walk, Run with Retrieval-Augmented Generation"!

I will use the same dataset through the three phases. It's British Columbia government's HR policy PDF documents. I've downloaded them all, bundled it in a tarball, and put them in a publicly readable GCS bucket access.

The Goal

This is mean to be educational and if you don't know how Retrieval-Augmented Generation works, this will hopefully get you a acquainted. If you do already know how it works, then maybe I can offer some new perspectives and tools that you can try out to enhance the results of you existing RAG system.

Crawl

The basics:

• Process PDFs
• Chunk + create embeddings
• Insert into local vector database
• Perform semantic search
• ADK agent to interact with the user

Walk

Builds on 'Crawl' phase:

• Improves document processing
• Improves chunking strategy
• Perform reranking after semantic search

Run

Run a fully-managed RAG system that applies the concepts covered in "Crawl" and "Walk" phases:

• Vertex AI RAG Engine
• Model Armor to provide guardrails