AI Features

KiBee uses AI at multiple layers to make guidance smarter over time — from understanding what a user wants in plain language to discovering recurring workflows that no one has written a flow for yet.

Intent matching

When a user asks KiBee a question via kibee.ask(), the intent engine finds the best matching flow without any manual keyword mapping. The base tier uses phrase scoring and page-context signals. Higher tiers add vector embedding similarity using pgvector for deeper natural-language understanding.

See Intent Matching for the full reference.

Workflow discovery

KiBee observes page-visit sequences across sessions and periodically clusters them into recurring journeys. The AI → Workflow recommendations view shows discovered patterns with suggested flow names, step sequences, and estimated user counts.

Discovered workflows become one-click flow drafts — review the suggestion, adjust any steps, and publish directly from the recommendations panel.

Journey embeddings

Each completed session is summarised as a journey text and embedded into a vector. These embeddings power:

• Semantic intent matching — finding flows by meaning, not keywords
• Journey clustering — grouping similar paths to find recurring patterns
• Recovery routing — returning stalled users to the exact step they were on

Embeddings are computed server-side using the OpenAI Embeddings API and stored in PostgreSQL via pgvector. No user data leaves your deployment unless you enable this feature explicitly.

Friction AI

The AI layer on friction data surfaces actionable recommendations rather than raw numbers. For each high-friction page it generates:

• A plain-language summary of where users are getting stuck
• Whether a guided flow or a product change is more likely to help
• A suggested flow draft when guidance is recommended

Inference outcomes

Every intent resolution and journey match is recorded as an inference outcome. Outcomes can be marked correct or incorrect from the Hive session detail, creating a feedback loop that improves matching over time.

Data and privacy

All AI features operate on anonymised session data. No PII is included in embeddings or model inputs unless you call kibee.identifyUser() and the user has consented to analytics. AI features can be disabled per workspace in Settings → AI.