Skip to main content
The Intent System classifies the operator’s request so Lucia can use the correct reasoning path and response shape.

​
Purpose

The intent system prevents Lucia from treating every message like a generic chat request. The current doctrine is: 
Human first. Purpose second. Boundary third.
It determines whether the operator is asking for: 
priority
deferral
arrival awareness
human reassurance
maintenance focus
system navigation
workspace orientation
next-step sequencing
payment reconciliation

​
Current Runtime Role

Intent classification supports Focus Ops behavior. It helps Lucia decide:
  • what mode to use
  • what data matters
  • what should be ignored
  • what response structure fits the moment
In v0.1.3.6, Focus Ops also uses 06 - Semantic Conversational Intent Assist for semantic lightweight utility and conversational intent classification. This layer rejects phrase patching. It classifies by prompt family and purpose, not by one lucky exact string. Current live-dev action-routing doctrine: 
CTA label is copy.
Structured action intent / metadata is routing truth.
Calendar booking clicks route to Full Booking Page record/review.
Structured Focus Ops actions route through Dynamic Action Workspace when work needs completion.
Intent classification should produce enough structured meaning for Lucia to route:
infinite CTA languageResolver Matrixfinite Action Workspace
If the operator names a guest, service, request, or issue, that scoped entity must be preserved through classification and routing. Payment reconciliation intent must preserve the difference between: 
payment truth review
booking note save
confirmed paid state
pending review state
actual payment mutation
Admin DAW payment reconciliation can save review notes, but payment status is not changed by that note. Payment mutation intent should not be invented because the current Admin UI has no charge, refund, mark-paid, or create-checkout operator control. If the operator arrives from Signal Stream Chat, active_context should scope the request before generic ranking. If the operator is working beside the app, active_context.workspace should orient workspace questions to the current surface. If the operator asks a short follow-up, prior recommendation memory should resolve the reference against the last verified concrete recommendation. If the operator gives a short confirmation such as “yes,” prior offer context should resolve it against the last offered action when verified. Explicit prompt subject still wins over page context. A named guest, service, request, or issue must not be overwritten by the current surface. Intent output must not leak raw implementation language to the operator: 
active_context
current_surface
payload
metadata

​
Example Intents

​
priority_triage

Used when the operator asks: 
What matters most right now?
Where should I start?
I am stretched thin.
Expected behavior: 
Pick the single highest-priority operational issue.

​
deferral

Used when the operator asks: 
What can wait until tomorrow?
What can I ignore right now?
Expected behavior: 
Identify safe-to-defer items without making the operator inspect everything.

​
human_utility

Used when the operator sends social, simple, or emotionally soft input. Example: 
Good morning
Thanks Lucia
I feel overwhelmed
Expected behavior: 
Respond naturally, briefly, and calmly.

​
semantic_conversational_assist

Used when the operator sends lightweight utility, social momentum, or scoped context prompts that should be understood without forcing a full operational triage. Example families: 
Time?
What time is it in Bangladesh?
Nice to see you
Let's do this
Should we keep dinner outside tomorrow?
Is outside still okay for the ceremony?
Expected behavior: 
Recognize the semantic family, answer or clarify within Lucia's operator boundary, and avoid drifting into open-domain chat.

​
maintenance_focus

Used when maintenance signals or property-ops issues are central. Expected behavior: 
Treat guest impact, arrival timing, departure timing, stay windows, payment/booking context, and task state as ranking signals.

​
Correct Intent Behavior

Lucia should:
  • classify quietly
  • route cleanly
  • protect semantic families by meaning
  • avoid exposing internal machinery unless useful
  • maintain stable response shape

​
Incorrect Intent Behavior

Lucia should not:
  • overexplain classification
  • phrase-patch individual prompts as a substitute for intent understanding
  • force every message into operational triage
  • respond emotionally when the user asked for action
  • respond operationally when the user needs containment
  • become open-domain ChatGPT

​
Truth-State Rule

Intent classification is not truth by itself. It is a routing decision. If the intent is uncertain, Lucia must behave conservatively and avoid overclaiming.

​
See Also