Intent Matching

Intent matching lets KiBee understand what a user wants in plain language and automatically select the best flow to run — no hard-coded triggers required.

How it works

When you call kibee.ask(query), KiBee:

Sends the query and the current page's visible targets to the API
Scores every flow for the current page against the query using phrase matching and page-context signals
Returns the best match with a confidence score between 0 and 1
The SDK can then run the matched flow automatically

Matching signals

Signal	Weight	Description
Phrase match	High	Exact or fuzzy match against a flow's `phrases` array.
Goal similarity	Medium	Semantic similarity between the query and the flow's `goal` text.
Visible targets	Medium	Target labels visible in the viewport increase the score for flows referencing those targets.
Page scope	Boost	Flows matching the current `pageId` are prioritised over flows from other pages.

Using intent matching

TypeScript

const match = await kibee.ask("how do I invite someone?");

if (match.flowId) {
  kibee.speak(match.message); // e.g. "I'll guide the invite team flow."
  await kibee.runInteractiveFlow(match.flowId);
} else {
  kibee.speak("I'm not sure how to help with that yet.");
}

IntentMatch response

Field	Type	Description
`flowId`	`string \| null`	The matched flow ID, or `null` if no match was confident enough.
`confidence`	`number`	Score from 0 to 1. Scores above ~0.5 are considered a match.
`message`	`string`	Human-readable explanation of the match, suitable to show in the bee's speech bubble.

Improving match quality

Add phrases — include all natural-language variations a user might type: “invite team”, “add a user”, “send an invite”.
Write a clear goal — the goal is used for semantic similarity. Write it as a complete sentence describing the outcome.
Use descriptive target labels — “Send invite button” is a stronger matching signal than “button-2”.
Scope flows to the right pageId — page-scoped flows are prioritised. If a flow only makes sense on the settings page, give it the settings page ID.

No AI costs by default. The base engine uses phrase scoring and page-context signals with no external LLM calls. Optional embedding-based semantic matching is available for deeper natural-language understanding.