Smart Connections — AI auto-linking
TL;DR — Smart Connections is a community plugin that builds embeddings for every note in your vault, then surfaces semantically similar notes as you read. Your Claude Code workflow is the active half (Claude reads + writes); Smart Connections is the passive half (always-on background recommendations).
What it does
Section titled “What it does”A pane in Obsidian that always shows “notes semantically related to this one.” Not by tag, not by link — by meaning.
Reading your weekly review and it surfaces a project log from 3 months ago that hits the same theme. You’d never have found it through search or backlinks.
Install
Section titled “Install”Settings → Community plugins → Browse → “Smart Connections” → Install + Enable.
First run: it builds embeddings for every note. Takes a minute for a small vault. Bigger vaults can take 10+ minutes — let it run in the background.
By default it uses local embeddings (no API key, no cost, no privacy concern). You can swap in OpenAI embeddings for better quality, but it’s optional.
The pane
Section titled “The pane”Open any note. The Smart Connections pane (right sidebar by default) lists related notes ranked by semantic distance. Click any to jump there.
You’ll discover:
- Same idea, different vocabulary. A Zettel about “biased anchoring” connects to a meeting note where someone said “they’re stuck on the first number they heard.” Tags wouldn’t have caught it.
- Forgotten threads. Old project notes that prefigured ideas you’re having now.
- Implicit themes. When the top 5 connections all share a vibe, you’ve found a theme worth naming.
Pairing with Claude Code
Section titled “Pairing with Claude Code”Two ways:
- Claude as the link-creator. Smart Connections suggests; Claude evaluates and writes the actual
[[link]]. Workflow: highlight a Smart Connections suggestion, ask Claude “is this connection real? If yes, link these two notes and add a 1-line explanation.” - Claude reading Smart Connections output. Future-state: a skill
auto-linkruns nightly, pulls top suggestions for newly-modified notes, asks Claude to vet each, and applies the high-confidence ones.
Privacy
Section titled “Privacy”If you use the local embeddings model (default), nothing leaves your machine. If you switch to OpenAI embeddings, your note text gets sent to OpenAI’s embedding API — fine if that’s your tolerance, off the table if not.
When NOT to use it
Section titled “When NOT to use it”- Vault under 100 notes. Manual linking is fine; nothing to discover.
- Privacy-paranoid. Even local embeddings store vectors. Some folks don’t want vault-derived data on disk in any form.
- Confused vault. If you can’t explain what most of your notes are for, more suggestions won’t help. Tidy first, automate after.