Global Rule Routing

Purpose

Global Rules should not be injected as a static block for every question. Claustrum routes rules dynamically so the bundle stays relevant while preserving critical safeguards.

Core Rules

• pinned=true rules are always included.
• severity=high rules are always included (subject to budget warnings).
• Other rules are selected by routing score.

Routing Modes

• semantic: token-similarity focused.
• keyword: direct keyword/tag overlap focused.
• hybrid (default): semantic + keyword blend.

Workspace settings:

• global_rules_routing_enabled (default: true)
• global_rules_routing_mode (default: hybrid)
• global_rules_routing_top_k (default: 5)
• global_rules_routing_min_score (default: 0.2)

Score Model

For non-pinned/non-high rules:

score = semantic_similarity * w_sem + keyword_overlap * w_kw + priority_weight + recency_weight - length_penalty

Where:

• semantic_similarity: cosine similarity over query and rule tokens.
• keyword_overlap: overlap ratio of query tokens vs rule content/tags.
• priority_weight: higher weight for lower numeric priority.
• recency_weight: newer rules get a small boost.
• length_penalty: very long rules are slightly penalized.

Query Source

Routing query (q_used) is selected in this order:

1. Explicit q from /v1/context/bundle
2. Pseudo-query built from project summary, active work, recent activity, and current subpath

Budget Interaction

Routing ranks rules first, then token budget is applied. If omitted rules become large, summary fallback is used per workspace settings.

Admin UI

Workspace → Global Rules:

• Routing enabled toggle
• Routing mode select
• top_k input
• min_score input

Project → Context Debug:

• Enter test query
• Load debug bundle
• Inspect selected/dropped rule IDs and score breakdown

Operational Guidance

• Keep top_k small (3~8) for better focus.
• Increase min_score when too many low-quality rules appear.
• Add tags (security, commit, naming, etc.) to improve keyword routing quality.