Tune Smart Connections scoring and ranking
Connections result shaping has three layers: candidate scope, primary scoring, and final ranking.
Score and ranking controls expose the two stages that shape final Connections order.
Choose Sources or Blocks
Select Smart Sources for note-level candidates or Smart Blocks for heading/block-level candidates.
Candidate selection happens before scoring. Fix an overly broad pool before tuning the score.
Configure candidate filters
Include and exclude path filters narrow the candidate pool before scoring and ranking are applied.
Filters decide which candidates are eligible.
Use include/exclude rules for project paths, archives, templates, tags, or other structural boundaries before adding a more complex scoring algorithm.
Choose a score algorithm
Score algorithms compute the primary relevance value.
| Algorithm | Use |
|---|---|
| Cosine Similarity | Stable baseline without feedback. |
| Similarity Adjusted by Feedback | Penalize candidates similar to hidden results. |
| Similarity Weighted by Feedback | Boost pinned-like signals and dampen hidden-like signals. |
| Similarity Weighted by Key + Frontmatter | Apply path, key, heading, or metadata multipliers. |
Choose a ranking algorithm
The ranking control separates final-order choices from the primary score algorithm.
Ranking algorithms reorder already-scored candidates.
| Algorithm | Use |
|---|---|
| None | Preserve score order. |
| Re-ranking model | Apply a configured Smart Rank model to top candidates. |
| Recency rank | Make modification time dominate final order. |
What is the difference between scoring and ranking?
Scoring decides the primary relevance value. Ranking changes the final order after scoring.
Use feedback-aware scoring
Feedback-aware algorithms use pinned and hidden signals.
Use them only when those actions represent stable preference. A one-off hide can otherwise create an unintended recurring bias.
Weight paths, keys, headings, and frontmatter
Weight configuration multiplies the base score when key fragments or metadata match.
{
"key_weights": {
"Projects/": 1.2,
"Readwise/": 0.8
},
"meta_weights": {
"status:evergreen": 1.15,
"type=spec": 1.1
}
}
Multiple matches multiply together.
Configure a reranking model
Re-ranking requires a configured Smart Rank model in Smart Environment Pro.
If no compatible ranking model is available, use None or Recency rank.
Rank by recency
Recency rank reorders the scored candidates by modification time.
Use it when freshness should dominate. Do not use it to solve an eligibility or semantic-scope problem.
Troubleshoot broad, weak, or noisy results
Apply tuning in this order:
- Fix candidate scope.
- Fix filters and limits.
- Establish Cosine Similarity as the baseline.
- Add score weighting.
- Add ranking only when the right candidates remain in the wrong order.
Why did changing three settings at once make results harder to evaluate?
Scope, score, and ranking changed simultaneously. Return to a baseline and alter one layer at a time.
Should I use Dedupe when similar notes keep appearing?
Use Dedupe when repeated material needs a keep, merge, archive, or ignore decision. Use algorithm tuning when the issue is result ordering.
Tune in the correct order
Choose Sources or Blocks, narrow the candidate pool, establish Cosine Similarity as the baseline, change scoring only when relevance is consistently wrong, and add ranking only when the right candidates exist in the wrong order.
Compare algorithm changes
Keep the anchor note, candidate pool, and filters stable while testing one change. Compare useful results across several representative notes rather than comparing score ranges alone.