Model routing

How ConfPlanner maps LLM roles to concrete models, and the keys that control the mapping. Code calls withLlmByRole("cheapest"|"best") (or withDefaultLlm()); application.yml binds each role to a model, so the same action can run on different models in different deployments without code changes. This page is the authoritative routing table.

Roles and the actions that use them

Role	Call	Default model	Used by
cheapest	withLlmByRole("cheapest")	gpt-4.1-nano	extractAttendeeProfile, shortlistSessions
best	withLlmByRole("best")	gpt-4.1	researchSessions, assembleSchedule
default	withDefaultLlm()	gpt-4.1-mini	premiumBriefing, planNetworking

loadCatalog and confirmSchedule use no model (plain code). No action declares a web tool group.

Configuration keys

Set in src/main/resources/application.yml under embabel.models:

Key	Description
embabel.models.default-llm	Model for withDefaultLlm() and the default role
embabel.models.llms.cheapest	Model bound to the cheapest role
embabel.models.llms.best	Model bound to the best role
embabel.models.llms.<name>	Bind any custom role name, then call withLlmByRole("<name>")

Swap any of these for a model your configured provider exposes — including a local Ollama tag for a role whose data must not leave the building.

Where it connects

• Set a provider/credentials: how to run with a real model.
• Assign models per role, hands-on: how to route models by role.
• Route a step to an on-prem model, hands-on: how to route a step to a local model.
• Why route by return-type complexity, and the data-residency trade-off: right-sizing models.
• The per-action rationale table also lives in the repo's MODEL_ROUTING.md.