Choosing a Selection Algorithm
This guide helps you select the right model selection algorithm for your use case.
Quick Decision Tree
Do you need deterministic routing?
├── Yes → Static Selection
└── No → Continue...
Do you have user feedback available?
├── Yes, abundant → Elo Rating
├── Some feedback → Hybrid (with Elo component)
└── No feedback → Continue...
Do you have good model descriptions?
├── Yes → RouterDC
└── No → Continue...
Is cost optimization important?
├── Yes → AutoMix
└── No → Static or Hybrid
Algorithm Comparison
Core Algorithms
| Algorithm | Feedback Needed | Setup Complexity | Adaptability | Cost Optimization |
|---|---|---|---|---|
| Static | None | Low | None | Manual |
| Elo | High | Medium | High | Indirect |
| RouterDC | None | Medium | Medium | No |
| AutoMix | Low | Medium | Medium | Yes |
| Hybrid | Varies | High | High | Yes |
RL-Driven Algorithms
| Algorithm | Feedback Needed | Setup Complexity | Adaptability | Personalization |
|---|---|---|---|---|
| Thompson | Medium | Low | High | Per-user option |
| GMTRouter | Medium | High | Very High | Built-in |
| Router-R1 | None | High | High | Via LLM |
Use Case Recommendations
Startup / MVP
Recommended: Static
Start simple with explicit rules. Migrate to adaptive methods as you collect data.
algorithm:
type: static
static:
default_model: gpt-3.5-turbo
High-Volume Production
Recommended: AutoMix or Hybrid
Optimize costs while maintaining quality at scale.
algorithm:
type: automix
automix:
cost_quality_tradeoff: 0.4
User-Facing Applications
Recommended: Elo
Let user feedback drive model selection for subjective quality.
algorithm:
type: elo
elo:
k_factor: 32
storage_path: /data/elo-ratings.json
Specialized Domains
Recommended: RouterDC
When models have distinct specializations, match queries to model capabilities.
algorithm:
type: router_dc
router_dc:
require_descriptions: true
Enterprise / Multi-objective
Recommended: Hybrid
Balance multiple factors: quality, cost, user satisfaction, and specialization.
algorithm:
type: hybrid
hybrid:
elo_weight: 0.3
router_dc_weight: 0.3
automix_weight: 0.2
cost_weight: 0.2
Personalized Multi-User Platforms
Recommended: Thompson Sampling or GMTRouter
Learn individual user preferences over time.
algorithm:
type: thompson
thompson:
per_user: true
min_samples: 10
Research / Complex Routing Logic
Recommended: Router-R1
When routing decisions require semantic understanding that's hard to encode in rules.
algorithm:
type: router_r1
router_r1:
router_endpoint: http://localhost:8001
use_cot: true
Migration Path
A typical progression as your system matures:
- Start: Static selection with simple rules
- Add feedback: Migrate to Elo as you collect user feedback
- Add descriptions: Add RouterDC for query-model matching
- Optimize cost: Incorporate AutoMix for cost efficiency
- Combine: Use Hybrid to leverage all methods
Key Considerations
Data Requirements
- Static: No data needed
- Elo: Needs consistent user feedback (thumbs up/down)
- RouterDC: Needs quality model descriptions
- AutoMix: Needs accurate pricing and quality scores
- Hybrid: Combination of above
- Thompson: Needs feedback; works online
- GMTRouter: Benefits from interaction history; can pre-train
- Router-R1: Needs router LLM server; model descriptions help
Latency Impact
| Algorithm | Typical Latency |
|---|---|
| Static | Under 1ms |
| Elo | Under 2ms |
| RouterDC | 2-5ms (embedding) |
| AutoMix | Under 3ms |
| Hybrid | 3-5ms |
| Thompson | Under 2ms |
| GMTRouter | 5-15ms (GNN) |
| Router-R1 | 100-500ms (LLM) |
Maintenance
- Static: Manual rule updates
- Elo: Self-maintaining with feedback
- RouterDC: Update descriptions when models change
- AutoMix: Update pricing when costs change
- Hybrid: Periodic weight tuning recommended