DeepSeek Multi-Model Router Rules

## DeepSeek Multi-Model Router Configuration

### Model Selection Matrix

| Task Type | Primary Model | Fallback | Notes |
|-----------|--------------|----------|-------|
| Simple Q&A | deepseek-chat (V3) | — | Fastest, cheapest |
| Code generation | deepseek-chat (V3) | claude-coder | V3 handles most coding well |
| Code review | deepseek-chat (V3) | deepseek-reasoner | Use R1 for security audits |
| Math/Logic | deepseek-reasoner (R1) | — | Always use R1 for math |
| Debugging | deepseek-reasoner (R1) | deepseek-chat | R1 excels at root cause analysis |
| Architecture | deepseek-reasoner (R1) | deepseek-chat | Complex tradeoff analysis |
| Creative writing | deepseek-chat (V3) | — | R1 is overly analytical |
| Data analysis | deepseek-chat (V3) | deepseek-reasoner | Use R1 for complex statistical reasoning |
| Translation | deepseek-chat (V3) | — | V3 handles 100+ languages |

### Routing Rules
1. Classify the task using keyword detection and intent analysis
2. Check complexity: if tokens > 2000 or nested sub-problems > 3, escalate to R1
3. Apply cost constraints: if budget remaining < 20%, prefer V3
4. Implement fallback: if primary model returns low-confidence response, retry with fallback

### Configuration
- V3 temperature: 0.7 (general), 0.3 (code), 0.9 (creative)
- R1 temperature: 0.6 (always)
- Max tokens: V3=4096, R1=8192
- Timeout: V3=30s, R1=120s

### Monitoring
Track model selection distribution, accuracy per model per task type, and cost per task type to continuously refine routing rules.