Add Semantic Routing or Mixture-of-Models skill to Emerging Techniques

Issue Description

Overview

Add a new skill for Semantic Routing or Mixture-of-Models (vLLM Semantic Router) to the 19-emerging-techniques category. Semantic Routing provides system-level intelligence for Mixture-of-Models (MoM) through signal-driven decision engine and plugin chain architecture for intelligent LLM routing, security, and optimization.

What is Semantic Routing?

Semantic Routing is an intelligent routing layer that uses signal-driven decisions and plugin chains to:

1. Route queries intelligently across multiple specialized models (math → Qwen-Math, code → DeepSeek-Coder)
2. Optimize costs by using smaller models for simple tasks, larger models for complex ones
3. Secure LLM systems with built-in jailbreak, PII, and hallucination detection
4. Reduce latency through semantic caching (10-100× speedup)
5. Enable model collaboration through Mixture-of-Models (MoM) architecture

Key Features

Signal-Driven Decision Engine:

• 10 signal types: keyword , embedding, domain/MMLU, fact_check, user_feedback, preference, language, latency (TPOT/TTFT), context, complexity
• Flexible combination: AND/OR operators for complex routing logic
• Multi-signal fusion: Combine signals for higher accuracy than single classifiers

Plugin Chain Architecture:

• semantic-cache - 10-100× latency reduction for similar queries
• jailbreak - Adversarial prompt detection and blocking
• pii - Personally identifiable information detection
• system_prompt - Dynamic system prompt injection per route
• header_mutation - HTTP header manipulation for routing control
• hallucination - Token-level hallucination detection during generation

Model Training: https://huggingface.co/llm-semantic-router

Why This Belongs in Emerging Techniques

1. Novel approach: System-level intelligence for MoM (vs. model-level MoE)
2. Production-ready: Used in real-world vLLM deployments
3. Research-backed: NeurIPS 2025 MLForSys paper, ICLR 2026 RouterArena #1 ranking
4. Cost-effective: 80-90% cost reduction vs. always using largest model
5. Active development: Regular releases, bi-weekly community meetings, AMD partnership

Proposed Skill Structure

19-emerging-techniques/semantic-routing/
├── SKILL.md                    # 200-500 lines main guidance
├── references/
│   ├── README.md              # Architecture overview
│   ├── signals.md             # 10 signal types deep dive
│   ├── plugins.md             # Plugin chain architecture
│   ├── training.md            # ModernBERT + LoRA training guide
│   ├── deployment.md          # Docker/Kubernetes deployment
│   ├── api.md                 # API reference
│   └── issues.md              # Common issues and solutions
└── examples/
    ├── basic-routing.yaml     # Simple keyword routing
    ├── multi-signal.yaml      # Complex signal combination
    └── production-stack.yaml  # Full production setup

Content Outline

SKILL.md (200-500 lines):

1. When to Use

   • Multi-model collaboration scenarios
   • Cost optimization needs
   • Security requirements (jailbreak/PII/hallucination)
   • Semantic caching for latency reduction

2. Quick Start

   pip install vllm-sr
   vllm-sr serve
   

3. Core Concepts

   • Mixture of Models (MoM) vs. Mixture of Experts (MoE)
   • Signal-Driven Decisions (10 signal types overview)
   • Plugin Chain Architecture (6 plugins overview)

4. Two Complete Workflows with Checklists

   • Workflow 1: Basic Multi-Model Routing

     • Define signals (keyword + domain)
     • Configure decision rules (AND/OR)
     • Set model mappings
     • Test routing decisions
     • Validate routing accuracy

   • Workflow 2: Production Deployment

     • Configure security plugins (jailbreak + PII)
     • Enable semantic cache
     • Set up monitoring metrics
     • Configure multiple backend models
     • Load testing
     • Deploy to Kubernetes

5. When to Use vs Alternatives

   • vs. LiteLLM (simple routing only)
   • vs. LangChain Router (slow LLM-based routing)
   • vs. Hand-written if-else (hard to maintain)

6. Common Issues

   • Signal conflicts resolution
   • Inaccurate routing decisions
   • High latency troubleshooting
   • Low cache hit rate optimization
   • Model loading failures

references/ (300KB+ target):

• signals.md: Detailed documentation of all 10 signal types with configuration examples, latency comparison, use cases, and combination strategies
• plugins.md: Deep dive into 6 plugins, plugin development guide, execution order
• training.md: Why ModernBERT, 4 classifier models, LoRA training methodology, datasets, performance metrics
• deployment.md: Docker Compose, Kubernetes + Helm, production configuration, performance tuning, observability
• api.md: OpenAI-compatible API, routing API, classification API, configuration API
• issues.md: Real GitHub issues, common errors and solutions, debugging methods

examples/:

• basic-routing.yaml: Simple keyword-based routing
• multi-signal.yaml: Multi-signal combination (keyword + domain + embedding)
• production-stack.yaml: Full production config with plugins, monitoring, multiple models

Key Highlights to Emphasize

Why Use Semantic Router?

• Cost optimization: Use Llama-3-8B for simple queries, GPT-4 for complex ones
• Quality improvement: Route math to Qwen-Math, code to DeepSeek-Coder
• Security built-in: Jailbreak, PII, hallucination detection
• Performance boost: 10-100× latency reduction via semantic cache

Core Advantages:

1. Multi-signal fusion: 10 signals combined > single classifier
2. Low latency: keyword 1ms, embedding 10-50ms, domain 50-100ms
3. Extensible: Plugin architecture for custom signals and processing
4. Production-ready: Kubernetes-native, Prometheus metrics, OpenTelemetry tracing

Resources

• GitHub: https://github.com/vllm-project/semantic-router (513 source files)
• Documentation: https://vllm-semantic-router.com (24,000+ lines)
• Paper: When to Reason: Semantic Router for vLLM (NeurIPS 2025)
• Blog: https://blog.vllm.ai/2025/09/11/semantic-router.html
• Community: vLLM Slack #semantic-router channel

Acceptance Criteria

• SKILL.md with proper YAML frontmatter (name: semantic-routing)
• 200-500 lines of focused guidance in SKILL.md
• 300KB+ reference documentation from official sources
• At least 2 complete workflows with checklists
• Code examples with language tags (yaml, bash, ```python)
• "When to use vs alternatives" section
• Common issues and solutions section
• References one level deep from SKILL.md (no nested references)
• Examples directory with 3 runnable configuration files

Related Skills

• 12-inference-serving/vllm - vLLM inference engine (backend for semantic router)
• 14-agents/langchain - Agent frameworks that can benefit from intelligent routing
• 15-rag - RAG systems that benefit from semantic caching and routing
• 16-prompt-engineering/dspy - Prompt optimization with routing decisions

---

Labels: enhancement, new-skill, emerging-techniques, documentation