Finetune examples are short sample conversations stored on a node. They steer the LLM through the prompt; no model is trained. There are two kinds:
| UI label | Field | Steers |
|---|---|---|
| Example replies | finetune_conversation_examples |
How the agent words its replies in this node |
| Transition examples | finetune_transition_examples |
Which node to move to next, given a conversation |
Add examples#
- Select a node and open Node settings at the bottom of its settings.
- Under Example replies, click Reply example. Each example starts with a User and an Agent turn. Click + Turn to add more and switch each turn between User and Agent.
- Under Transition examples, click Transition example, pick the node to Go to, and write the turns that should lead there.
The fields exist on every node type.
Fields#
{
"finetune_conversation_examples": [
{
"id": "ex_1a2b3c4d",
"transcript": [
{ "role": "user", "content": "How much is a check-up?" },
{ "role": "agent", "content": "A standard check-up and clean is $195. Would you like to book one?" }
]
}
],
"finetune_transition_examples": [
{
"id": "ex_5e6f7a8b",
"transcript": [{ "role": "user", "content": "Actually can someone call me back about my bill?" }],
"destination_node_id": "node_billing"
}
]
}| Field | Type | Description |
|---|---|---|
id |
string | Unique per example |
transcript |
array of {role: "user" or "agent", content} |
The example turns |
destination_node_id |
string | Transition examples only: the node this conversation should lead to |
How they are used#
Example replies are added to the system prompt while the node is active, under "Example conversations" with the line "Respond in the style of these examples." Every reply in the node sees them. They are rendered with {{variables}}.
Transition examples are added under "Transition examples" to the separate LLM call that decides prompt transitions, each labelled with the transition_to_<node_id> function it should lead to. That separate call is used:
- on conversation and subagent nodes that have an else edge, and
- on every non-conversation node (function, code, MCP, extract and so on) that has prompt transitions.
On a conversation node without an else edge, prompt transitions are chosen inside the reply call, where only the example replies are included. To make transition examples count on such a node, give it an else edge.
Writing good examples#
- Keep each example to the few turns that matter. Long examples cost tokens on every turn in the node.
- Show the edge cases: the caller who answers two questions at once, the off-topic question, the caller who says "no" to the offer.
- Keep examples consistent with the node's instruction, so the prompt does not give the LLM two different answers.