Vault Intelligence

Appearance

GARS Search and Context Tuning Guide ​

This document explains the technical parameters in the Search and context tuning section of the Advanced Settings. These values govern the Graph-Augmented Relevance Score (GARS) and the Context Assembly logic.

1. Search expansion ​

Expansion trigger threshold (Default: 0.70) ​

Determines which search results are "strong" enough to trigger a graph neighbor expansion.

  • How it works: If the top search result has a score of 10.0, any result with a score >= 7.0 (70%) will have its neighbors (linked notes, topic siblings) pulled into the candidate pool.
  • Tuning: Lowering this (eg to 0.50) creates a broader, more exploratory search. Raising it (eg to 0.90) makes the search tighter and more focused on direct matches.

Expansion seeds limit (Default: 5) ​

A safety cap on how many documents can trigger expansion.

  • How it works: Even if 20 documents pass the threshold, only the top 5 will actually have their neighbors expanded.
  • Tuning: Increase for very dense repositories where relationships are more important than text. Decrease if background indexing performance is a concern.

Absolute expansion floor (Default: 0.40) ​

The absolute minimum score required for any expansion.

  • How it works: If your search results are weak (vague queries), the system will not expand neighbors for results scoring below this floor, preventing "fishing" for irrelevant connections.

2. Context assembly (the accordion) ​

The system uses Relative Relevance to decide how much of a note to show the AI. This is calculated as a percentage of the top match's score.

Primary threshold (Default: 0.90) ​

  • Action: Full file body is included.
  • Logic: These are high-confidence matches that are nearly as relevant as the top result.

Supporting threshold (Default: 0.70) ​

  • Action: Extracts contextual snippets around query terms.
  • Logic: These notes provide useful background or supporting evidence but aren't the main answer.

Structural threshold (Default: 0.35) ​

  • Action: Displays only the note structure (headers) as a "Table of Contents".
  • Logic: These are peripheral notes or graph neighbors. They provide structural context without bloating the prompt.
  • Cap: This mode is strictly capped at the top 10 matches to prevent metadata noise.

3. Advanced weights ​

Spreading activation weight (Default: 0.25) ​

Determines how much "bonus" score a graph neighbor receives from its parent match.

  • Logic: If Note A matches the query, its neighbor Note B receives 25% of Note A's score automatically.
  • Tuning: Increase to 0.50 to make the AI much more aware of "connectedness". Decrease to 0.10 if you want the AI to stay strictly within the search results.

Neighbor decay (Default: 0.30) ​

The penalty applied as you move further away in the graph.

  • Logic: Every hop away from a direct search match reduces the activation bonus by this factor.
  • Tuning: Leave at 0.30 for calibrated behavior in most vaults.

4. Domination prevention ​

Single doc soft limit (Default: 0.10) ​

Prevents a single large "Primary" document from accidentally crowding out other relevant results in the context window.

  • Logic: Even if a document is a 100% match, it is "soft-capped" at 10% of the total context budget if other relevant documents are available. This ensures the AI always has a diverse set of sources to draw from.