Skip to content
Basalt

Connection (C)

Element: Carbon (C · 6) · forms more bonds than any other element.

What it does

Finds pairs of notes whose embedded content is highly similar, while their explicit link graph is at distance ≥ 2 — i.e. you wrote the same thing twice and never connected the two.

Algorithm sketch

  1. For each note, find its k-nearest neighbors in embedding space (k=20)
  2. Filter to neighbors at graph distance ≥ 2 (i.e. no shared link/backlink)
  3. Apply hub-penalty so daily-notes don’t dominate
  4. Sort by similarity × novelty (embedding similarity × inverse hub-frequency)

False positives

  • Templated notes. Two notes from the same template will look similar. Add the template path to .basalt/config.toml::ignored_templates.
  • Quoted material. If two notes quote the same paper, they’ll match. The quote extractor in the parser handles this for top-k cases.

Confidence

Connection confidence is the raw cosine similarity (0–1). Above 0.78 is a real match; below 0.65 is noise.