Recall over neatness
Broad result packs include snippets, ranks, channels, and scores so an agent can inspect more context instead of trusting one brittle hit.
Local-first Obsidian recall
Obsidian HighRecall
Find the notes your AI agent or normal search would otherwise miss.
A recall-first search wrapper for large private Obsidian vaults. It reuses Smart Connections vectors when available, falls back to local hybrid/fulltext search, and returns broad context packs for Codex or CLI workflows.
Local-first
Raw notes and private queries stay on the user's machine.
Smart + OHS
Uses Smart Connections vectors, OHS fallback, or both.
16 tasks
Private-vault benchmark reports anonymized aggregate metrics.
Public fixture
Install and evaluator behavior can be tested without a private vault.
Try it without sharing a vault.
The public fixture checks install, syntax, query execution, and recall metrics before anyone points the tool at private notes.
git clone https://github.com/ToussaintKnight/obsidian-high-recall-skill.git
cd obsidian-high-recall-skill
npm test
node skills/obsidian-high-recall/scripts/obsidian_high_recall.mjs query \
"data collection for embodied AI robot demonstrations" \
--vault docs/fixtures/demo-vault \
--backend smart \
--limit 10Built for recall-heavy research memory.
Precision can be cleaned up downstream. A missed relevant note is harder to recover.
Private by default
The workflow reads local Markdown and writes derived runtime data outside the vault. Public artifacts contain aggregate/anonymized data only.
Reproducible doorway
A small public vault and smoke test make the tool testable before users run it on a real knowledge base.
The private boundary is explicit.
Codex, the vault, Smart vectors, OHS fallback, merge/rank, recall packs, and private evaluation stay local. GitHub receives only the portable skill, docs, and aggregate figures.
Benchmark signal, not a universal claim.
The current private-vault study uses 16 manually labeled recall tasks and reports anonymized aggregate metrics. It is meant to compare deployment choices, not to claim universal Obsidian search quality.
| Condition | Precision@20 | Recall@20 | F1@20 | MRR | Mean latency |
|---|---|---|---|---|---|
| Smart | 0.17 | 0.57 | 0.26 | 0.65 | 1.00s |
| OHS | 0.17 | 0.55 | 0.25 | 0.24 | 49.54s |
| RRF union | 0.18 | 0.61 | 0.28 | 0.61 | 50.54s |
Test it on a real vault.
Run the fixture first, then try one broad query on your own vault. Report OS, Node version, Obsidian setup, Smart Connections status, and whether it found notes you would have missed.