A developer with 52,000 markdown files in Obsidian shows that you don't need vector databases or RAG pipelines to give LLMs rich context. Files are nodes. Wikilinks are edges. Folders are taxonomy. That's the whole architecture.
The setup uses a PARA-inspired structure with /projects/, /areas/, /people/, and /daily/ directories. An LLM agent handles the grunt work: creating meeting notes, pulling Google Docs, linking everything to relevant person and project files. After a 1:1 with their boss, the agent spins up a note with wikilinks to that person and whatever projects came up. Over months, each person's note becomes a running timeline of every conversation and decision.
The real insight is treating this as a context engineering system, not just a wiki. Breaking big coding projects into small requirements is a popular alternative to relying on vector search. When you need to write a design doc, you point the LLM at a project folder containing six months of meeting notes, prior docs, and Slack threads. The output beats cold prompting because the LLM works with actual project history instead of your summary of it. The approach came from Andrej Karpathy's recent posts about using LLMs to build personal knowledge bases.
Technically this swaps probabilistic semantic retrieval for deterministic structural retrieval. Vector databases use embeddings and cosine similarity to find relevant chunks. The file-system graph uses explicit wikilink relationships and folder hierarchies. You skip chunking strategies and indexing entirely.
Hacker News commenters raised privacy concerns about cloud models and noted that LLMs organize files well but degrade quality when allowed to write content directly into the knowledge base. Trust issues with agents, such as hallucinating commands, have also been highlighted.
The unsolved problem is automated inbox processing. The author describes a "graveyard of experiments" trying to define what "processed" means. Every attempt has been too rigid or too loose. But getting started doesn't require 52,000 files. Create the folder structure, have an agent create and link meeting notes for a week, then point it at those folders next time you need to draft something.