A Personal RAG System: Turning Obsidian Notes Into An AI Reference Library

When people hear RAG, they often think of vector databases first.

Embeddings. Chunks. Retrieval. Context injection.

But for personal knowledge management, that is often the wrong starting point. The retrieval system is not the main problem. The real problem is that the knowledge being retrieved is not yet clean enough to be useful.

For me, the practical starting point is Obsidian. I already have notes I have read, rewritten, linked, and used. A personal RAG system begins when those notes become a reference library for an AI.

RAG Is An Open-Book Exam

Retrieval-Augmented Generation means the model looks something up before answering.

In a closed-book exam, the model answers from its parameters.
In an open-book exam, it uses documents, notes, manuals, policies, or code supplied at query time.

For personal knowledge work, the real question is:

If the AI does not know my thinking,
what reference library should I build for it?

The issue is not only whether the model is smart. The issue is whether I have left knowledge in a form that can be read again.

Do Not Put The Inbox Into RAG

The first mistake is to index the entire vault.

An Inbox contains rough ideas, duplicated clippings, unfinished thoughts, and source fragments that have not yet become your own knowledge. If those notes go directly into retrieval, the AI cannot easily tell judgment from noise.

Start with reviewed notes instead.

Good RAG candidates have a few properties.

• The title has a clear scope.
• Source material and personal interpretation are separated.
• One note answers one governing question.
• Frontmatter records domain, status, and public-use boundaries.
• The note is linked to related notes.

RAG begins with note hygiene.

Frontmatter Becomes The Usage Contract

Humans can infer context. Agents need explicit signals.

---
domain: "blog-ops"
status: "reviewed"
agent_usable: true
publishable: true
trust: "medium"
source_type: "personal-note"
updated: 2026-06-24
---

This is small, but it changes the system.

agent_usable says whether the note may be used.
publishable says whether it can become public material.
trust records confidence.
source_type separates source, summary, and personal interpretation.

As a RAG system grows, metadata becomes as important as retrieval accuracy.

Start With Long Context Before Infrastructure

You do not need a vector database on day one.

For a small personal system, collect a set of reviewed Markdown notes, place them into a long-context model, and ask concrete questions.

Based on this note bundle,
extract the five educational principles I repeat most often,
and map them to possible blog categories.

The goal is not to build the final system. The goal is to see whether your notes can be read by an AI at all. If the answer is vague, the retrieval system may not be the problem. The note structure may be vague.

Retrieval Must Be Verified By Sources

RAG quality should not be judged only by whether the answer sounds good.

The better question is: what notes did the answer rely on?

Check:

• Did the cited note actually exist?
• Did the answer preserve the note’s meaning?
• Did it show conflicting notes when they mattered?
• Did it treat old notes as current?
• Did it use private notes in a public answer?

Without this layer, RAG does not remove hallucination. It merely gives hallucination a source-shaped costume.

LLM Wiki And RAG Are Complementary

I do not want to discard RAG.

I want the order to be right.

First, build the LLM Wiki: names, scopes, relationships, status, sources, and permissions.

Then add RAG: retrieve the right piece of that organized knowledge at the right time.

The LLM Wiki is the library classification system.
RAG is the librarian.

Without classification, the librarian can still run very fast, but may bring back the wrong book.

The Point

Do not teach the AI from scratch every time.

Let it refer to the knowledge you have already built.

That requires notes that are good for humans and usable by agents: clear title, metadata, links, sources, status, and publication boundaries.

RAG does not begin with a vector database. It begins when your knowledge can be read again.