Retell Conversation Flows: Branches, Knowledge and Call Review

A single prompt asks a language model to remember the whole call procedure. A structured flow makes the path visible: greeting, question, branch, action and close. That helps when a call needs predictable steps such as checking eligibility before offering an appointment.

Make branches testable

Give each branch a condition you can verify from a caller’s answer. For example, route “yes, I’m the property owner” to the booking step, “no” to a polite close, and an uncertain answer to one clarification. A function or action node should define what data it needs, what happens on success and what the agent says if the action fails. Test the branch boundaries, not only the happy path.

Attach knowledge where it is useful

A knowledge source should answer a defined class of questions, such as pricing or service coverage. Keep documents current and separate unrelated policies so the agent retrieves the right material. If the platform lets you scope sources to steps, use that scope to reduce irrelevant matches. Ask the same question in several ways and check whether the answer cites the intended policy rather than an outdated or neighboring document.

Use post-call fields as a review queue

Choose a small set of fields before you build: outcome, booking status, callback request and any required qualification answer. Validate those fields against transcripts. Treat an extracted value as a useful signal, not unquestionable ground truth, especially when the caller used an ambiguous phrase.

Retell’s product interface and feature set evolve, so verify current node and analytics behavior in the official Retell documentation. Keep test calls and expected results alongside the flow definition. When a call fails, you can inspect the node, source document or action responsible instead of rewriting the entire prompt.

Continue with the LeadOro AI Phone Agent Stack series for the related walkthrough.