Build your own RAG (Retrieval Augmented Generation) agent in 25 minutes. If you're building AI products, or you want to be, you've heard the term thrown around. We believe in learning by doing, so on this episode we teach you how to build your own RAG agent from scratch. You'll learn key terminology like vector store and embedding, and you'll have a working agent by the end. Walk away with the confidence to talk about RAG with your business and technical stakeholders.

The workflow examples from this episode are available for download on Github here. Simply open a new workflow, click the import from URL button, and paste the link from Github.

A step-by-step guide can be found here.

Chapters

00:00 - What Is RAG and Why Product Teams Should Care
04:10 - Tools and Prerequisites for the Build
07:07 - Building the Data Ingestion Workflow in N8N
13:11 - Connecting Embeddings and Document Loaders
17:20 - Building the Chat Agent
21:50 - Testing the RAG Agent Live

Key Topics

RAG (Retrieval Augmented Generation): How RAG lets an LLM search over specific documents instead of pulling from its entire training data
Vector Databases: What they are, how they store information for LLM retrieval, and why Supabase works well for this
Embeddings Models: How Cohere's embedding model translates text into a format LLMs use for similarity search
N8N Workflow Setup: Step-by-step walkthrough of building both the data ingestion and chat agent workflows
Dimension Matching: Why your embeddings model and database table must use the same number of dimensions or your results will be useless
The Think Tool: How a scratchpad tool helps AI agents remember why they made decisions during multi-step processes
Metadata in Vector Stores: Adding properties like author, likes, and retweets to give the LLM more context about stored documents