Type: explanation · Last reviewed: 2026-07-12
Applies to: Pulse Copilot · edge/on-prem or managed cloud
A plant question like "why is idfan_vibration rising on boiler-2?" is really several steps: find the right tags, pull their history, run the numbers, weigh them against limits and past events, and write it up. Pulse Copilot does those steps for you. Behind it is Clarity AI — an agentic multi-agent system (built on Google's ADK) in which one orchestrator delegates each step to a specialized sub-agent and streams the answer back.
This page explains how the layer is organized and why you can trust it. For the catalogue of questions it answers by role, see Copilot use cases.
The root pulse_manager never does the computation itself — it classifies intent and delegates:
system_health_agent.incident_agent.Each sub-agent has one job. meta_data_agent is the only one that fetches data; data_analysis_agent only analyzes what was already fetched (inside a RestrictedPython sandbox); the incident/system agents diagnose; dashboard_creation_agent recommends from five visualization types (lines, histogram, scatter, table, comparison); question_agent proposes what to investigate next.
The single most important design rule of the layer: every factual statement about an incident or system must come from a tool response in the current turn. The orchestrator has no pre-loaded plant knowledge to fall back on, so it cannot quietly invent a number or a cause — if the data isn't fetched, it says so. This anti-hallucination discipline is what makes a plain-English answer safe to act on.
"Why is idfan_vibration rising on boiler-2?" becomes a chain of delegations:
pulse_manager → classify: incident/diagnosis (L4), resolve unit boiler-2
→ meta_data_agent : list incidents, fetch idfan_vibration history
→ incident_agent : examine deviating tags, plot, add recurrence context
→ data_analysis_agent: compute supporting statistics
pulse_manager → stream a written diagnosis + charts back to the user
Everything in the reply traces to those tool calls. The role-by-role list of what people actually ask is in Copilot use cases.
Deployment note: the layer runs in the AI bundled (edge) topology (bundled with the historian install, air-gap capable) or AI separate (cloud) (a managed service pointed at an on-prem historian). A Historian-only install runs no AI service. In the edge-bundled build (
CLARITY_APP), the system-health and incident-summary tools are excluded — data retrieval, analysis, dashboards, and questions remain. See Overview § Technical topologies.
Sessions persist in SQLite (edge) or PostgreSQL (cloud), and answers stream to the browser token-by-token over SSE.
agent · sub-agent · tool · ADK · session · RestrictedPython sandbox · Clarity AI