Tiers and Agents may use¶
Tiers¶
A tier is how strong you say a model is: fast, balanced or strongest. You set it in the Tier column of a provider's Models tab. Fill tiers fills them from the catalogue.
Tiers let KOLONİ pick a model for a job instead of one fixed name:
- a role can ask for a tier instead of one model;
- an agent can call a job easy or hard, to go one tier down or up;
- a chat's Orchestrator can follow how hard your message is.
A model with no tier is only ever picked by name.
The tier lists¶
Below the provider cards, the Models page shows three lists: Strongest, Balanced and Fast. Each lists its models, numbered in the order they are tried. Drag a model, or use its move up and move down buttons, to change the order; each move saves at once. A model that agents may not use is marked "off for agents". With no tiers at all, the page says "No model has a tier yet".
A model greyed with its provider's state ("not answering") shows how that provider looked when the page read it. A provider that is not installed, not signed in or without a key is passed over until KOLONİ sees it again: the page reads it, or you press its Check. A model server is passed over after not answering for two minutes. Where every model of a tier looks so, the first is still tried.
Agents may use¶
Each model has an Agents may use switch. On, agents may pick the model by themselves, for example by its tier. Off, only you pick it: in a chat's model picker, or as a role's own model.
A provider with no list of models has one switch for the whole provider: Agents may use this provider. Never run a model that is not on this list keeps agents to the listed models.
What it is good for (for a provider) and Its note for agents (for a model) are short notes that agents read when they choose.
A role's model¶
On the Agents page each role chooses who picks its model: Orchestrator, As the chat, One tier (Fast, Balanced or Strongest) or One model. See Roles.
Orchestrator model: automatic¶
In the chat's model picker, Orchestrator model: automatic sends each message to a fast model, which judges it easy, normal or hard. The Orchestrator then runs on a fast, balanced or strongest model. That is one more request for each message you send. It overrides the Orchestrator role's own model and tier. A trigger's message does not use it.