scripts/config/server_config.json

{
  "servers": [
    {
      "name": "creative",
      "url": "http://localhost:1234/v1",
      "model": "mistral-7b-instruct",
      "timeout": 120.0,
      "max_concurrent": 4,
      "weight": 1.0,
      "enabled": true,
      "description": "Fast, creative model for research tasks"
    },
    {
      "name": "analytical", 
      "url": "http://localhost:1235/v1",
      "model": "llama-3.1-8b-instruct",
      "timeout": 120.0,
      "max_concurrent": 4,
      "weight": 1.0,
      "enabled": true,
      "description": "Balanced model for analysis and critique"
    },
    {
      "name": "synthesis",
      "url": "http://localhost:1236/v1", 
      "model": "mixtral-8x7b-instruct",
      "timeout": 180.0,
      "max_concurrent": 2,
      "weight": 1.5,
      "enabled": true,
      "description": "High-quality model for synthesis and final output"
    },
    {
      "name": "fallback",
      "url": "http://localhost:1237/v1",
      "model": "phi-3-mini-instruct",
      "timeout": 60.0,
      "max_concurrent": 8,
      "weight": 0.5,
      "enabled": false,
      "description": "Fast fallback model for high-load scenarios"
    }
  ],
  "agent_mapping": {
    "research": "creative",
    "analysis": "analytical", 
    "synthesis": "synthesis",
    "critic": "analytical"
  },
  "load_balance_strategy": "agent_type_mapping",
  "health_check_interval": 30.0,
  "failover_enabled": true,
  "debug_mode": false
}