Note Relatedness Signals

Note relatedness signals reference

The signals notes_link_discovery.py scores when it looks for related-but-unlinked pairs among your personal notes (the corpus outside AgentMemory/ + .obsidian/). The audit never mutates a personal note — it surfaces candidate links for operator review (A3). Suggestions are strictly personal↔personal; an AgentMemory/ entry is never a source or a target (DC-2).

⚡ Quick Reference

Question	Answer
What runs the audit?	harness/skills/memory/scripts/notes_link_discovery.py (the relatedness engine + report writer).
How do I see suggestions?	python3 harness/skills/memory/scripts/notes_link_discovery.py --format text (or --format json).
Which notes are in the corpus?	Personal notes only — the Obsidian vault excluding AgentMemory/, .obsidian/, .trash, .git (DC-2).
What are the two signals?	TF-IDF content overlap (lexical) + embedding cosine (semantic, opt-in via --embeddings). Folder + date proximity are weak context.
Does the audit ever edit a note?	Not by default — read-only / surface-only (DC-1). The opt-in --apply flag is the one exception: it writes the safe suggestions into a marked ## Related section, backup-first + idempotent (you directed it; A3 satisfied).
Where do outputs live?	Report → <vault>/_meta/notes-links-<date>.md; embedding cache → <vault>/_meta/notes-embeddings.json. Both under the agent-controlled vault, never beside a personal note, never the AgentMemory vec-index.db.
How do I run the report?	See Find missing note links.
Related pages	Find missing note links

Signals

The personal-notes corpus has no usable graph signal — only a handful of ~390 notes carry tags, one has a [[wikilink]], and frontmatter is just title / created / updated. So tags, links, and frontmatter fields are dead signals here; relatedness is content-based.

Signal	Role	Notes
TF-IDF term overlap (title + body)	Primary (always on)	Lowercase alphanumeric tokens ≥3 chars; English + Spanish stopwords stripped; HTML/CSS/#hex/image-embed/URL markup stripped before tokenizing (many notes are clipped web content); title terms double-weighted; sublinear-tf × IDF, L2-normalized, cosine-scored via an inverted index.
Embedding cosine (title + lede)	Secondary (opt-in --embeddings)	Each note embedded with the local BGE model (embed.py), full pairwise cosine. Catches same-topic/person/event pairs with little shared vocabulary — including cross-language pairs TF-IDF structurally can't see. Graceful-skips to TF-IDF-only when sentence-transformers is absent.
Folder context	Secondary	Same-folder pairs are labelled same folder in the report (navigational grouping); cross-folder pairs show A ⇄ B.
Date proximity	Available	created / updated are parsed into the corpus but not scored in v1.
Existing [[wikilinks]]	Exclusion only	A pair already linked (by stem or relative path, either direction) is never re-suggested.
Tags	Dead	~no notes carry tags — no signal in this corpus.
Frontmatter fields	Dead	Only title / created / updated present — no signal.

Thresholds

Flag	Default	Meaning
--min-score	0.18	TF-IDF cosine floor. Pairs below it are dropped.
--embed-min-score	0.70	Embedding cosine floor (BGE similarity runs hot — ~0.3–0.5 even for unrelated prose — so the semantic threshold sits well above the lexical one).
--top	40	Per-signal shortlist cap (0 = all). Keeps the report a high-signal shortlist, not noise.

The corpus is ~390 notes (≈76k candidate pairs); the TF-IDF inverted index compares only term-sharing notes, a document-frequency band drops per-note noise and ubiquitous boilerplate, and the embedding pass is a bounded O(n²) cosine over cached vectors. The report deduplicates: a pair found by both signals appears once (in the TF-IDF section, flagged ✓ also semantically related); the embedding section lists only the pairs TF-IDF missed.

Related

• Find missing note links — the operator recipe that runs the audit and reads the report.