HarkHarkDocs
Theme

Lookups, actions, and AI nodes

The nodes that do work: look up customer context, take actions like creating a ticket, and run AI steps that degrade honestly.

Beyond talking to the customer, a flow can look things up, take real actions, and call on AI. These are the nodes that let a flow resolve an issue rather than just route it.

Lookup nodes

A lookup fetches context and branches on whether it found anything (a found path and a not-found path). The available lookups are:

  • Identify customer: resolve who the visitor is.
  • Open conversations: find the visitor's existing open conversations, so the flow can offer to continue one instead of starting fresh.
  • Recent orders (placeholder): a stub for order lookups, ready for a future data source.

A lookup can save its result under a variable for later nodes to use.

Action nodes

An action performs a real change and branches on success or failure. The actions are:

ActionWhat it does
Create ticketOpens a support ticket; saves the ticket id.
Create taskCreates a work item; saves the task key.
Add tagApplies a tag to the conversation.
Set prioritySets the conversation's priority.
GitHub: create issueOpens a GitHub issue; saves the issue URL.
GitHub: commentComments on a linked GitHub issue.
Slack: notifySends a Slack notification.
Continue by emailMoves the conversation to email so it can carry on asynchronously.

Action parameters accept variable substitution, so a created ticket can carry the subject the customer typed earlier.

AI nodes

An AI node runs a language-model step. There is one node type with a kind for each job:

  • Classify intent: label what the customer wants.
  • Search knowledge: pull relevant context from your public corpus (help articles, feedback, changelog).
  • Generate answer: draft a reply, optionally grounded in your knowledge.
  • Summarize: condense the conversation so far.
  • Detect sentiment and detect urgency: read the emotional tone or urgency.
  • Extract: pull structured fields out of free text.

Every AI node routes three ways: done (a usable result at or above a confidence threshold you set), low confidence (a result below the threshold), and failure (the model was unavailable or errored). This is the honest-degrade principle in action: when the AI cannot help, the flow takes the failure branch and hands off to a human rather than inventing an answer. Routing on an AI result stays composable: a classify-intent node saves its label to a variable, and a plain condition node branches on it.