LiyaEngine

modelConfig controls which LLM models are used for inference and how cost is managed per request and session.

Schema

{
  "modelConfig": {
    "strategy": "hybrid",
    "default_model": "gpt-4o-mini",
    "premium_model": "gpt-4o",
    "tool_calling_model": "gpt-4o",
    "intent_model_overrides": {
      "mock_interview": "gpt-4o",
      "candidate_scoring": "gpt-4o"
    },
    "max_tokens_per_request": 4000,
    "max_cost_per_request": 0.10,
    "max_cost_per_session": 1.00,
    "max_cost_per_day": 100.00,
    "max_context_tokens": 4000,
    "max_session_messages": 20,
    "session_timeout_minutes": 30
  }
}

Model strategy

`strategy`

Value	Behaviour
`hybrid`	Standard intents use `default_model`; premium intents use `premium_model` (default)
`all-mini`	All intents use `default_model` — lowest cost
`all-standard`	All intents use `premium_model` — highest quality

Premium intents (default)

In hybrid mode, these intents use premium_model by default:

mock_interview
coaching_session
candidate_prescreen
candidate_scoring

Override the premium intent list via featureConfig.premium_intents.

`intent_model_overrides`

Pin a specific model to a specific intent, overriding strategy:

{
  "intent_model_overrides": {
    "resume_analysis": "gpt-4o-mini",
    "career_transition_planning": "gpt-4o"
  }
}

Cost limits

Field	Default	Description
`max_cost_per_request`	`$0.10`	Hard cap per API call — request rejected if estimated cost exceeds this
`max_cost_per_session`	`$1.00`	Session-level cap — new messages rejected when session cost hits this
`max_cost_per_day`	`$100.00`	Daily tenant-level cap
`max_tokens_per_request`	`4000`	Output token limit per request

Context limits

Field	Default	Description
`max_context_tokens`	`4000`	Token budget for session history injected into context
`max_session_messages`	`20`	Maximum messages carried in session memory
`session_timeout_minutes`	`30`	Session expires after this period of inactivity

When you've connected your own provider credentials, model routing defers to your configured provider with automatic fallback if it fails. The strategy, default_model, and premium_model fields still control which model names are requested — but they're sent to your provider, not Liya's. See Models + Router.

Model Configuration