Conversation node

The Conversation node (type: "conversation") is where the agent talks with the caller. Most flows are mostly conversation nodes connected by transitions, with automatic nodes in between.

Instruction#

Pick Prompt or Static Sentence on the node card or under Instruction in Node settings.

Both render {{variables}}. When the agent speaks first, a static sentence on the start node is the greeting.

Write prompt instructions as a task for this step only:

Ask for the caller's full name and the day they would like to come in.
If they give a day without a time, ask morning or afternoon.
Do not offer specific times yet.

Transitions#

New conversation nodes start with one empty prompt transition. Add more with + on the card (Prompt, Equation, Else, Always/Skip) or Add under Transitions in Node settings.

What happens on a caller turn#

Without an else edge:

1. The always edge, then equation edges, are checked before the LLM runs. If one matches, the flow moves and the next node answers.
2. Otherwise one LLM call runs with the node's instruction and the prompt transitions (plus global nodes) as functions. The LLM either replies or calls a transition; on a transition, the next node answers in the same turn.

With an else edge:

1. Always and equation edges are checked.
2. A separate LLM call decides among the prompt transitions and global nodes.
3. If none applies, the else edge is taken.
4. The node you land on generates the reply.

An else edge makes the node leave on every caller turn (if it is connected), at the cost of one extra LLM call per turn. Use it for nodes that should only ever ask one question.

Skip response#

A skip response edge fires after the agent finishes speaking in the node. Use it for one-way statements such as disclosures: a static sentence node with a skip response edge says its line and moves on without waiting for the caller.

Skip with contact memory#

Skip with contact memory (skippable: true) lets the agent skip the node when what it collects is already known about the caller. It needs Use memory in calls turned on under Agent Settings, Contact memory ("Adds {{contact_memory}} to the prompt and lets flow nodes marked skippable be skipped."), and a caller with stored memory.

When the flow enters a skippable node, one LLM call checks whether the caller's memory fully answers what the node's instruction collects. If so, the node is skipped once per call and the flow follows the node's first connected edge (or its else edge when there are no edges).

Tools#

Conversation nodes do not get the flow's functions. To let the LLM call functions freely while talking, use a subagent node. To call a function at a fixed point, use a function node.

Fields#

{
  "id": "node_collect",
  "type": "conversation",
  "name": "Collect details",
  "instruction": { "type": "prompt", "text": "Ask for the caller's name and preferred day and time." },
  "edges": [
    { "id": "edge_1", "destination_node_id": "node_extract", "transition_condition": { "type": "prompt", "prompt": "The caller gave their name and a preferred day and time" } }
  ]
}

Gotchas#

• A node with no matching transition keeps talking. If the caller goes off script, the node's instruction should tell the agent how to steer back.
• Static sentences are spoken exactly as written on entry. Use a prompt when the wording should adapt to what the caller said.
• Knowledge base retrieval runs on caller turns in every node, using the node's knowledge bases when it overrides them.

Related#