Multi-Model Architectures
Modern AI applications often use multiple models: specialized models for specific tasks, cost optimization with smaller models, and different providers for capabilities. ECM Protocol supports context routing across models.
Router Specification
Route Request
{
"operation": "route.context",
"task": {
"type": "summarization",
"input_tokens": 5000,
"required_capabilities": ["long_context"]
},
"context": {
"context_ids": ["ctx-1", "ctx-2"]
},
"constraints": {
"max_latency_ms": 2000,
"max_cost_per_1k_tokens": 0.01
}
}
Route Response
{
"selected_model": {
"provider": "anthropic",
"model": "claude-3-haiku",
"estimated_latency_ms": 800,
"estimated_cost": 0.005
},
"context_strategy": {
"injection_point": "system",
"format": "xml",
"truncation": "relevance"
}
}
Model Registry
Model Capabilities
Registry tracks model capabilities:
{
"model_id": "claude-3-opus",
"capabilities": {
"context_window": 200000,
"supported_tasks": ["reasoning", "coding", "analysis"],
"languages": ["en", "fr", "de", ...],
"vision": true,
"function_calling": true
},
"performance": {
"avg_latency_ms": 2000,
"throughput_tpm": 100000
},
"cost": {
"input_per_1k": 0.015,
"output_per_1k": 0.075
}
}
Routing Strategies
Capability-Based
Route based on required capabilities. Match task requirements to model capabilities. Filter to capable models first.
Cost-Optimized
Minimize cost while meeting requirements. Use cheaper models when sufficient. Reserve expensive models for complex tasks.
Latency-Optimized
Minimize response time. Consider model latency characteristics. Account for context preparation time.
Quality-Optimized
Maximize output quality. Route to best model for task type. Benchmark-informed routing.
Fallback Protocol
Cascading Fallback
Define fallback chains for failures. Primary model fails → try secondary. Include rate limit handling. Circuit breakers for unhealthy models.
Context Adaptation
Adapt context for fallback models. Truncate for smaller context windows. Reformat for different requirements. Preserve essential information.
Conclusion
Multi-model routing enables optimal model selection for each task. The ECM Protocol routing specification provides standardized model registry, routing rules, and fallback handling.