The semantic layer problem in PKM

Someone in a Reddit thread asked about my auto-linking setup — whether the links catch structural connections or just vocabulary overlaps. The honest answer is mostly vocabulary. Abstract ideas can connect to almost anything, which makes it hard to judge whether a link is useful or just plausible.

• Auto-links surface things that share obvious words and miss the pairs where the connection is structural — same shape of problem, different subject.
• Improving that without touching thousands of notes individually feels like another full-time job. That's the line that resonated most in the thread.
• The answer probably isn't better tooling for manual linking — it's improving the semantic layer, how the system understands relationships rather than vocabulary. The person I was talking to runs Obsidian with embeddings and an LLM on top. We both landed in the same place: manual linking isn't the bottleneck.
• Database theory might offer useful frameworks here — the semantic layer is essentially a schema problem.
• For project-specific linking, a simpler approach: tag notes with a job ID and filter on that when writing. You only need the links when working on something specific.

The question isn't "how do I link better" but "how do I make the system link better without me." I don't have the answer yet, but it's not about reviewing notes by hand.

With zero orphan notes,
Zhenia