Domain Based Routing

This guide shows you how to use fine-tuned classification models for intelligent routing based on academic and professional domains. Domain routing uses specialized models (ModernBERT, Qwen3-Embedding, EmbeddingGemma) with LoRA adapters to classify queries into categories like math, physics, law, business, and more.

Key Advantages

• Efficient: Fine-tuned models with LoRA adapters provide fast inference (5-20ms) with high accuracy
• Specialized: Multiple model options (ModernBERT for English, Qwen3 for multilingual/long-context, Gemma for small footprint)
• Multi-task: LoRA enables running multiple classification tasks (domain + PII + jailbreak) with shared base model
• Cost-effective: Lower latency than LLM-based classification, no API costs

What Problem Does It Solve?

Generic classification approaches struggle with domain-specific terminology and nuanced differences between academic/professional fields. Domain routing provides:

• Accurate domain detection: Fine-tuned models distinguish between math, physics, chemistry, law, business, etc.
• Multi-task efficiency: LoRA adapters enable simultaneous domain classification, PII detection, and jailbreak detection with one base model pass
• Long-context support: Qwen3-Embedding handles up to 32K tokens (vs ModernBERT's 8K limit)
• Multilingual routing: Qwen3 trained on 100+ languages, ModernBERT optimized for English
• Resource optimization: Expensive reasoning only enabled for domains that benefit (math, physics, chemistry)

When to Use

• Educational platforms with diverse subject areas (STEM, humanities, social sciences)
• Professional services requiring domain expertise (legal, medical, financial)
• Enterprise knowledge bases spanning multiple departments
• Research assistance tools needing academic domain awareness
• Multi-domain products where classification accuracy is critical

Configuration

Configure the domain classifier in your config.yaml:

classifier:
  category_model:
    model_id: "models/category_classifier_modernbert-base_model"
    use_modernbert: true
    threshold: 0.6
    use_cpu: true
    category_mapping_path: "models/category_classifier_modernbert-base_model/category_mapping.json"
  
  pii_model:
    model_id: "models/pii_classifier_modernbert-base_presidio_token_model"
    use_modernbert: true
    threshold: 0.7
    use_cpu: true
    pii_mapping_path: "models/pii_classifier_modernbert-base_presidio_token_model/pii_type_mapping.json"

categories:
  - name: math
    system_prompt: "You are a mathematics expert. Provide step-by-step solutions."
    model_scores:
      - model: qwen3
        score: 1.0
        use_reasoning: true
  
  - name: physics
    system_prompt: "You are a physics expert with deep understanding of physical laws."
    model_scores:
      - model: qwen3
        score: 0.7
        use_reasoning: true
  
  - name: computer science
    system_prompt: "You are a computer science expert with knowledge of algorithms and data structures."
    model_scores:
      - model: qwen3
        score: 0.6
        use_reasoning: false
  
  - name: business
    system_prompt: "You are a senior business consultant and strategic advisor."
    model_scores:
      - model: qwen3
        score: 0.7
        use_reasoning: false
  
  - name: health
    system_prompt: "You are a health and medical information expert."
    semantic_cache_enabled: true
    semantic_cache_similarity_threshold: 0.95
    model_scores:
      - model: qwen3
        score: 0.5
        use_reasoning: false
  
  - name: law
    system_prompt: "You are a knowledgeable legal expert."
    model_scores:
      - model: qwen3
        score: 0.4
        use_reasoning: false

default_model: qwen3

Supported Domains

Academic: math, physics, chemistry, biology, computer science, engineering

Professional: business, law, economics, health, psychology

General: philosophy, history, other

Features

• PII Detection: Automatically detects and handles sensitive information
• Semantic Caching: Cache similar queries for faster responses
• Reasoning Control: Enable/disable reasoning per domain
• Custom Thresholds: Adjust cache sensitivity per category

Example Requests

# Math query (reasoning enabled)
curl -X POST http://localhost:8801/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MoM",
    "messages": [{"role": "user", "content": "Solve: x^2 + 5x + 6 = 0"}]
  }'

# Business query (reasoning disabled)
curl -X POST http://localhost:8801/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MoM",
    "messages": [{"role": "user", "content": "What is a SWOT analysis?"}]
  }'

# Health query (high cache threshold)
curl -X POST http://localhost:8801/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "MoM",
    "messages": [{"role": "user", "content": "What are symptoms of diabetes?"}]
  }'

Real-World Use Cases

1. Multi-Task Classification with LoRA (Efficient)

Problem: Need domain classification + PII detection + jailbreak detection on every request Solution: LoRA adapters run all 3 tasks with one base model pass instead of 3 separate models Impact: 3x faster than running 3 full models, <1% parameter overhead per task

2. Long Document Analysis (Specialized - Qwen3)

Problem: Research papers and legal documents exceed 8K token limit of ModernBERT Solution: Qwen3-Embedding supports up to 32K tokens without truncation Impact: Accurate classification on full documents, no information loss from truncation

3. Multilingual Education Platform (Specialized - Qwen3)

Problem: Students ask questions in 100+ languages, ModernBERT limited to English Solution: Qwen3-Embedding trained on 100+ languages handles multilingual routing Impact: Single model serves global users, consistent quality across languages

4. Edge Deployment (Specialized - Gemma)

Problem: Mobile/IoT devices can't run large classification models Solution: EmbeddingGemma-300M with Matryoshka embeddings (128-768 dims) Impact: 5x smaller model, runs on edge devices with <100MB memory

5. STEM Tutoring Platform (Efficient Reasoning Control)

Problem: Math/physics need reasoning, but history/literature don't Solution: Domain classifier routes STEM → reasoning models, humanities → fast models Impact: 2x better STEM accuracy, 60% cost savings on non-STEM queries

Domain-Specific Optimizations

STEM Domains (Reasoning Enabled)

- name: math
  use_reasoning: true  # Step-by-step solutions
  score: 1.0           # Highest priority
- name: physics
  use_reasoning: true  # Derivations and proofs
  score: 0.7
- name: chemistry
  use_reasoning: true  # Reaction mechanisms
  score: 0.6

Professional Domains (PII + Caching)

- name: health
  semantic_cache_enabled: true
  semantic_cache_similarity_threshold: 0.95  # Very strict
  pii_detection_enabled: true
- name: law
  score: 0.4  # Conservative routing
  pii_detection_enabled: true

General Domains (Fast + Cached)

- name: business
  use_reasoning: false  # Fast responses
  score: 0.7
- name: other
  semantic_cache_similarity_threshold: 0.75  # Relaxed
  score: 0.7

Performance Characteristics

Domain	Reasoning	Cache Threshold	Avg Latency	Use Case
Math	✅	0.85	2-5s	Step-by-step solutions
Physics	✅	0.85	2-5s	Derivations
Chemistry	✅	0.85	2-5s	Mechanisms
Health	❌	0.95	500ms	Safety-critical
Law	❌	0.85	500ms	Compliance
Business	❌	0.80	300ms	Fast insights
Other	❌	0.75	200ms	General queries

Cost Optimization Strategy

1. Reasoning Budget: Enable only for STEM (30% of queries) → 60% cost reduction
2. Caching Strategy: High threshold for sensitive domains → 70% hit rate
3. Model Selection: Lower scores for low-value domains → cheaper models
4. PII Detection: Only for health/law → reduced processing overhead

Reference

See bert_classification.yaml for complete configuration.