This video discusses Andrej Karpathy's post about using Obsidian as a Retrieval-Augmented Generation (RAG) system for AI coding agents. The presenter highlights that traditional RAG systems can be complex to set up, whereas Obsidian offers a simpler, often sufficient solution for many users, especially those not dealing with millions of documents. A key advantage mentioned is that Obsidian is free. The core of this approach lies in Obsidian's file structure within a 'vault'. Raw information is stored in a 'raw' folder. Then, a large language model (LLM) processes this raw data and transforms it into interconnected 'wikis' within the vault. These wikis are designed to be queryable by both humans and the LLM, functioning similarly to a knowledge graph but on a simpler scale. The presenter also mentions having a detailed tutorial on their YouTube channel about setting this up.

Chase AI

So, this post by Andre Carpathy talking about how he uses Obsidian to essentially act as a RAG system with his agent coding agents is getting a ton of traction, and that's because normal RAG systems can be kind of difficult to set up.

But with something like Obsidian, we essentially get a 90% solution that is just as good for most people in most use cases if they aren't scaling into millions of documents. Plus, Obsidian is completely free, which is even better.

The real secret to how it works is in the file structure. We have everything inside of the vault. The vault is a folder designated by Obsidian where Obsidian actually goes to work. We then have a raw folder. That's where all of the essentially raw knowledge gets sent. And then we have a series of wiki folders, which is where the large language model comes in, and it takes all that information from the raw folder and turns it into different wikis essentially that you can query, and the large language model can query. And in essence, this is similar to something like a knowledge graph that you see here from Light RAG, just on a simpler scale.

So, I did a full deep dive on my YouTube of how to set this up.

Karpathy's Obsidian "RAG"

So, this post by Andre Carpathy talking
about how he uses Obsidian to
essentially act as a rag system with his
agent coding agents is getting a ton of
traction and that's because normal rag
systems can be kind of difficult to set
up. But with something like Obsidian, we
essentially get a 90% solution that is
just as good for most people in most use
cases if they aren't scaling into
millions of documents. Plus, Obsidian is
completely free, which is even better.
The real secret to how it works is in
the file structure. We have everything
inside of the vault. The vault is a
folder designated by Obsidian where
Obsidian actually goes to work. We then
have a raw folder. That's where all of
the essentially raw knowledge gets sent.
And then we have a series of wiki
folders which is where the large
language model comes in and it takes all
that information from the raw folder and
turns it into different wiks essentially
that you can query and the large
language model can query. And in
essence, this is similar to something
like a knowledge graph that you see here
from light rag just on a simpler scale.
So, I did a full deep dive on my YouTube
of how to set this up.

Get Our Mobile App

Karpathy's Obsidian "RAG"

Transcription