My RAG setup is just Obsidian now

Last February I built my first RAG setup for my own work and personal stuff. It worked well, and I kept improving it through the spring. Then I changed jobs and stopped having time to maintain it. Within a couple of months it was out of date enough that I stopped trusting the answers.

The setup I actually use now is a lot simpler. My notes are in Obsidian, which stores everything as regular markdown files in a folder, and I point the AI at the folder. I'm not running embeddings or a vector database anymore. For a few hundred notes, the model can find what it needs from the files themselves.

Steph Ango, who runs Obsidian, wrote a piece a while back called file over app. His argument is that your knowledge should live in plain files you own, so it outlasts whatever tool you happen to be using. I'd add that plain files turned out to be the most AI-compatible format I could have picked. Markdown needs no export step or parsing layer, and the links between my notes give the AI a map of how my thinking connects.

Keeping custom GPTs current was a chore, and I needed multiple of them. I was able to consolidate with a RAG setup, but it was still pretty intensive getting the proper documentation into the system, then working through correct chunking and the other common RAG headaches.

That's all gone with this new setup. The AI works with the vault the same way I do. When I ask something, it searches the folder, opens the notes that match, and follows the links between notes to pull in related context. A new note is part of the system the moment I save it. There's no separate knowledge base to feed, so the only maintenance is the note-taking I was already doing.

None of this would hold up at work, where the data is way too big for this approach and shared across teams with permissions attached, so we run real retrieval infrastructure there. But my personal vault is a few hundred files that only I use, and for that, plain files and a capable model are plenty.