Two questions about conversation-flow transitions, for predicting and
testing routing:
Are prompt transition conditions evaluated in a separate LLM call after
the user speaks, or in the same request that generates the reply (e.g.
edges exposed as functions)? Which model and settings are used for that
decision, and does it see the full call history or only the current
node’s turns?
Global nodes don’t seem to trigger from inside a component/subflow: a
user saying “I want a human” jumps to the global node from the main
flow, but inside a component the component’s else edge exits instead.
Is that intended, and is there a setting to allow global nodes inside
components?
It runs as a separate LLM call after the user’s turn, not as function/tool edges on the main reply model. Your agent LLM generates the reply; a dedicated small routing model decides whether to transition.
The routing model is a Retell-hosted small model.
It’s not the same as your agent LLM and it’s not configurable per agent.
It sees: the flow’s global instruction, the current node only (type + instruction), the list of candidate transition conditions (local edges on the current node plus eligible global-node conditions), any fine-tune examples on those edges, and the last 10–15 utterances of the transcript — not the full call history.
thank you for the extensive reply!
The first part actually answered both questions in a way. The reason it happens is because the se the separate llm model either mis-categorizes or misjudges the callers i.e. a user saying I want a human is looking for a next node transition that doesn’t exist when inside a component, thus leading to jumping to the global node.
Would be cool if that could be introduced as a feature someday that the model that’s responsible for the node transitions can also be selected and trained!
Most of the issues we’ve seen actually come from incorrect node transitions because a transition condition can’t always be either yes or no but are sometimes blurry without a clear line leading to the model having difficulty correctly assessing the situation and then transitioning to an incorrect node.
Thanks for the detailed follow-up and for sharing the additional context.
That feedback around node-transition ambiguity is helpful, especially for cases where transition conditions are not strictly binary and the model has to interpret a more nuanced user intent.
I’ve noted your suggestion about having more control over the model responsible for transition decisions. Thanks again for taking the time to explain the behavior you’re seeing so clearly.