I've been obsessed with a question lately: what if your LLM actually knew things?
Not retrieved things. Not hallucinated things. Known things — the way you know your own apartment layout, or your project's weird quirks, or which APIs break on Tuesdays.
The standard answer is RAG (retrieval-augmented generation). You embed your documents, you retrieve chunks on-demand, you inject context. It works. But there's something fundamentally off about it.
The Problem with RAG
RAG treats knowledge like a filing cabinet. You query it, you get a drawer, you read the drawer. The model never absorbs anything — it's always reaching.
The result: answers that feel assembled rather than understood. The model knows what the document says. It doesn't know what you meant when you wrote it at 2am with three browser tabs open and a half-formed intuition about the architecture.
Wikis Are Different
A wiki isn't a filing cabinet. A wiki is a mind map with timestamps.
Good wikis have:
- Bidirectional links — ideas know each other
- Edit history — you can see how understanding evolved
- Density gradients — some pages are stubs, some are dense, and that signal matters
- Author voice — you can tell which pages someone cared about
These are exactly the things RAG ignores. Chunk size is uniform. Retrieval is cosine similarity. The shape of the knowledge graph is invisible.
What I'm Building
The idea: treat an Obsidian vault as a first-class knowledge layer for an LLM, not just a document store.
Concretely:
- Parse the vault graph (links, backlinks, tags, frontmatter)
- Build a weighted context selector — not by similarity but by relevance topology
- Feed the model not chunks but knowledge paths — the chain of links between the query and the answer
query: "why did we switch from prisma to drizzle?"
→ find: /decisions/db-migration.md
→ traverse: backlinks, tagged #architecture
→ path: db-migration → performance-notes → prisma-limitations
→ context: three connected notes, ordered by reasoning flow
The model gets a story, not a pile of chunks.
Why This Matters
This isn't just a retrieval improvement. It's a different theory of what "memory" means for an AI assistant.
RAG says: memory is lookup. Wiki-native says: memory is structure.
The difference shows up in output quality. A model with structure can reason forward. It can notice contradictions between your old and new notes. It can tell you when your decisions.md hasn't been updated since you changed direction three weeks ago.
That's not retrieval. That's understanding.
More to come as I build this out. The Obsidian vault this site runs on is the first test case.