Skip to main content

​
Basic example

Select only fast models: 
on_http_request:
  - type: ai-gateway
    config:
      model_selection:
        strategy:
          - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_p95 < 1200)"
          - "ai.models.filter(m, m.metrics.global.error_rate.total < 0.01)"

​
How it works

Strategies execute in order: 
1. Filter to models with p95 latency < 1200ms
2. From those, filter to models with < 1% error rate
3. Select from the remaining models
If a strategy returns no models, the gateway tries the next strategy.

​
Performance-based filtering

​
Low latency

Prefer the fastest models: 
model_selection:
  strategy:
    - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_avg < 500)"
    - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_avg < 1000)"
    - "ai.models.random()"

​
High reliability

Prefer models with low error rates: 
model_selection:
  strategy:
    - "ai.models.filter(m, m.metrics.global.error_rate.total < 0.01 && m.metrics.global.error_rate.timeout < 0.005)"
    - "ai.models.random()"

​
Balanced

Balance speed and reliability: 
model_selection:
  strategy:
    - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_p95 < 2000 && m.metrics.global.error_rate.total < 0.02)"
    - "ai.models.random()"

​
Cost optimization

​
Prefer cheaper models

Select mini/turbo models when available: 
model_selection:
  strategy:
    - "ai.models.filter(m, m.id.contains('mini') || m.id.contains('turbo'))"
    - "ai.models.random()"

​
Cost tiers

Define cost tiers with fallback: 
model_selection:
  strategy:
    # Try budget models first
    - "ai.models.filter(m, m.id.contains('3.5-turbo'))"
    # Fall back to standard models
    - "ai.models.filter(m, m.id.contains('4o-mini'))"
    # Fall back to any model
    - "ai.models"

​
Provider preference

​
Prefer specific provider

Try a specific provider first: 
model_selection:
  strategy:
    - "ai.models.filter(m, m.provider_id == 'openai')"
    - "ai.models.random()"

​
Avoid specific provider

Exclude a provider unless necessary: 
model_selection:
  strategy:
    # Try everything except Google first
    - "ai.models.filter(m, m.provider_id != 'google')"
    # Fall back to Google if needed
    - "ai.models.filter(m, m.provider_id == 'google')"

​
Multi-criteria filtering

​
Complex performance criteria

model_selection:
  strategy:
    # Must be fast AND reliable
    - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_p95 < 1500 && m.metrics.global.error_rate.total < 0.01)"
    - "ai.models.random()"

​
Tiered selection

Prioritize different criteria: 
model_selection:
  strategy:
    # Tier 1: Fast and reliable
    - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_p95 < 1000 && m.metrics.global.error_rate.total < 0.005)"
    # Tier 2: Moderate speed, good reliability
    - "ai.models.filter(m, m.metrics.global.latency.upstream_ms_p95 < 2000 && m.metrics.global.error_rate.total < 0.01)"
    # Tier 3: Any model
    - "ai.models"

​
Metadata-based filtering

Use custom metadata to categorize and filter models: 
providers:
  - id: openai
    models:
      - id: "gpt-4o"
        metadata:
          tier: "premium"
          compliance: "hipaa"
      - id: "gpt-4o-mini"
        metadata:
          tier: "budget"
          compliance: "standard"
Strategy: 
model_selection:
  strategy:
    # Use budget tier for cost savings
    - "ai.models.filter(m, m.metadata.tier == 'budget')"
    # Fall back to any tier
    - "ai.models"

​
Compliance filtering

Filter by compliance requirements: 
model_selection:
  strategy:
    - "ai.models.filter(m, m.metadata.compliance == 'hipaa')"
    - "ai.models"

​
Feature-based filtering

Filter by supported capabilities: 
model_selection:
  strategy:
    # Only models that support tool calling
    - "ai.models.filter(m, 'tool-calling' in m.supported_features)"
    - "ai.models"

​
Vision models

Select models that can process images: 
model_selection:
  strategy:
    - "ai.models.filter(m, 'image' in m.input_modalities)"
    - "ai.models"

​
Real-world examples

​
High-volume production

Optimize for cost and speed: 
on_http_request:
  - type: ai-gateway
    config:
      providers:
        - id: openai
        - id: anthropic
        - id: google
      
      model_selection:
        strategy:
          # Prefer cheap, fast models with good reliability
          - "ai.models.filter(m, m.id.contains('turbo') || m.id.contains('mini')).filter(m, m.metrics.global.latency.upstream_ms_avg < 800 && m.metrics.global.error_rate.total < 0.02)"
          - "ai.models.random()"

​
High-reliability application

Optimize for reliability over cost: 
model_selection:
  strategy:
    # Only the most reliable models, prefer premium
    - "ai.models.filter(m, m.metrics.global.error_rate.total < 0.005 && m.metrics.global.error_rate.timeout < 0.001).filter(m, !m.id.contains('mini'))"
    - "ai.models.random()"

​
Development environment

Prefer self-hosted for development: 
model_selection:
  strategy:
    # Use self-hosted Ollama
    - "ai.models.filter(m, m.provider_id == 'ollama')"
    # Fall back to cloud for testing
    - "ai.models.filter(m, m.provider_id == 'openai' && m.id.contains('mini'))"
    - "ai.models"

​
Monitoring strategy effectiveness

Track which models are selected: 
request_count{model="gpt-4o"}: 45,234
request_count{model="gpt-4o-mini"}: 123,456
request_count{model="claude-3-5-sonnet"}: 5,432

latency_avg{model="gpt-4o"}: 1100ms
latency_avg{model="gpt-4o-mini"}: 650ms
Use metrics to refine your strategy.

​
Available variables

VariableTypeDescription
m.idstringModel identifier
m.provider_idstringProvider identifier
m.author_idstringModel author identifier
m.display_namestringHuman-readable model name
m.custombooleanWhether this is a custom model
m.metadataobjectCustom metadata
m.input_modalitieslistSupported input types
m.output_modalitieslistSupported output types
m.supported_featureslistSupported capabilities
m.metrics.global.request_countnumberTotal requests
m.metrics.global.latency.upstream_ms_avgnumberAverage latency (ms)
m.metrics.global.latency.upstream_ms_p95numberP95 latency (ms)
m.metrics.global.error_rate.totalnumberOverall error rate (0-1)
m.metrics.global.error_rate.timeoutnumberTimeout error rate (0-1)
m.metrics.global.error_rate.rate_limitnumberRate limit error rate (0-1)

​
Best practices

  1. Start simple - Begin with basic filters, add complexity as needed
  2. Include fallbacks - Always have a final strategy that accepts any model
  3. Monitor metrics - Use metrics to validate your strategy is working
  4. Test strategies - Test with production traffic patterns
  5. Use request_count - Filter by m.metrics.global.request_count > 0 to ensure metrics exist

​
See also