| name | mr-rug |
| description | MR.RUG (Mixture of Reasoning + Agentic GraphRAG) framework for deploying specialized reasoning experts who build unified knowledge graphs through Reliability-Aware RAG. Use when tasks require multi-expert collaboration, complex analysis needing 3+ perspectives, research synthesis, technical system design, or any R≥4 complexity task. Triggers on expert deployment, knowledge graph construction, RA-RAG retrieval, multi-domain synthesis, or when KTG-DIRECTIVE v28 Phase 1 is invoked. |
MR.RUG Framework v28
Phase 1 of KTG-DIRECTIVE | Expert deployment + Knowledge graph construction
Deploy specialized reasoning experts who build a unified knowledge graph through Reliability-Aware RAG, then semantically embed that graph into their cognitive architecture.
Architecture: M.R.R.U.G
| Phase | Component | Action |
|---|
| M | Mixture | Deploy experts based on mode |
| R | Role | Assign domains + handoff routes |
| R | RAG | RA-RAG retrieval with ARQ filtering |
| U | Update | Merge mini-graphs, resolve conflicts |
| G | Generate | Semantic embedding + expert embodiment |
Expert Deployment
By Mode:
- QUICK (R≤3): 0 experts (direct answer)
- ANALYTICAL (R=4-6): 3 experts (core specialists)
- DELIBERATE (R≥7): 5 experts (full team)
- MAXIMUM (R≥9): 5-8 experts (model-dependent)
Expert Profile Template:
Name: [Role identifier]
Domain: [Area of expertise]
Responsibility: [Accountable deliverables]
ARQ Standards: [Quality criteria for domain]
Tools: [Available resources]
Success Metric: [Measurable outcome]
Expert Archetypes
See references/archetypes.md for complete expert patterns:
- Technical (Architect, Implementation, Security, Performance, DevOps)
- Analytical (Domain, Research, Data Science, Strategic, QA)
- Creative (Director, Execution, Editor, Audience, Innovation)
- Cross-Domain (Integration, Translation, Synthesis)
Role Assignment
Execution Patterns:
Baton Passing (Sequential):
Expert-1: Node-1 → Node-3 → Node-5
Expert-2: Node-2 → Node-4
Handoff: Expert-1 output → Expert-2 input
Expert Swarm (Parallel):
Self-assignment based on specialty
Dynamic collaboration on complex nodes
No fixed routes
Permissions:
- Read: Full task context
- Write: Assigned nodes only
- Consult: Other experts
- Enrich: Parallel commentary
- Cannot: Override domains without consensus, skip ARQ gates, proceed with confidence <9/10
RA-RAG Protocol
Traditional RAG: Search → Retrieve → Summarize → Use
RA-RAG: Search → Retrieve → ARQ Filter → Categorize → Graph → Embody
Per Expert:
- Specialized Retrieval - Domain-specific queries
- ARQ During Collection - Quality introspection while retrieving
- Reliability Scoring - Source Authority, Recency, Relevance, Actionability (keep ≥7/10)
- Graph Categorization - Concepts, Facts, Relationships, Actions, Validations
See references/ra-rag-protocol.md for detailed retrieval patterns.
Knowledge Graph Construction
Node Types:
- CONCEPT: Core ideas, definitions
- FACT: Verified data points
- RELATIONSHIP: How concepts connect
- ACTION: Executable steps
- VALIDATION: Checks, constraints
Edge Types: REQUIRES, ENABLES, CONFLICTS_WITH, SUPPORTS, DEPENDS_ON
Merge Protocol:
- Collect mini-graphs from all experts
- Identify overlapping nodes (same concept, different sources)
- Resolve conflicts via source comparison
- Strengthen unanimous patterns
- Fill gaps from strongest sources
Semantic Embedding (Generate Phase)
Post-graph construction, each expert:
- Internalizes unified graph structure
- Extracts reusable reasoning patterns → BoT (if R≥6)
- Arms domain-specific ARQ quality gates
- Becomes ready for execution with full knowledge
Verbose Mode
Trigger /ktg-verbose for execution trace:
=== MR.RUG EXECUTION TRACE ===
[M] MIXTURE DEPLOYMENT: Mode, experts deployed
[R] ROLE ASSIGNMENT: Expert→Node mappings
[R] RA-RAG RETRIEVAL: Queries, docs, filter stats, graph size
[U] GRAPH MERGE: Total nodes/edges, conflicts resolved, gaps filled
[G] SEMANTIC EMBEDDING: Patterns extracted, ARQ armed
=== MR.RUG COMPLETE ===
Best Practices
DO:
- Match expert count to complexity
- Use RA-RAG for domains needing current knowledge
- Let experts build independent mini-graphs first
- Resolve conflicts with source citations
- Embed ARQ into expert cognition
- Save patterns to BoT (if R≥6)
DON'T:
- Deploy too many experts for simple tasks
- Skip RA-RAG phase
- Force premature consensus
- Ignore graph gaps
- Proceed without ARQ activation
Timing
| Phase | Duration |
|---|
| Expert deployment | Instant |
| RA-RAG retrieval | 30-60s per expert (parallel) |
| Graph synthesis | 10-20s |
| Semantic embedding | 5-10s |
| Total MR.RUG | ~60-90s (3-5 experts) |
Downstream savings: Structure planning 50% faster, execution 30% faster, verification 40% faster.
MR.RUG v28 | Part of KTG-DIRECTIVE v28 | Kevin Tan (ktg.one)