Skip to main content
Every night, toorow refreshes the data, runs its checks, and composes what is worth saying about the day. The result is a daily insight: a narrative that names the metrics that moved, the findings that fired, and the context the business registered — each one traceable to the pull it came from.

​
1. The nightly run

The nightly scheduler (SCHEDULER_ENABLED=true) triggers an automated evaluation loop at 02:00 local time:
  1. Daily Ingestion & Mart Build: Executes datastream pulls and runs dbt models to update canonical marts.
  2. Quality & Anomaly Checks: Evaluates data freshness, volume checks, and z-score anomaly detectors.
  3. Composition: Assembles the KPI deltas, the anomalies that fired, and the active business context into a daily insight.
An agent asks get_daily_insight_readiness(project_id) whether there is anything to say, composes the narrative with preview_daily_insight(project_id, payload), and publishes it with publish_daily_insights(project_id, insight_date). A published narrative is withdrawn with retract_daily_insight — an audited transition, never a deletion.

​
2. Business Context Events

A metric rarely moves on its own. A team registers context events — a redesign, a price change, a campaign launch, an outage — and the platform keeps them beside the numbers. When an agent analyzes metric anomalies, toorow correlates performance shifts against registered context events, providing instant explanatory context (e.g., “Conversions dropped 25% on July 14, coinciding with a scheduled checkout gateway migration”).

​
3. FastMCP Dual-Channel Response Model

Every MCP data tool in toorow enforces a dual-channel output model:
  • Text Channel: A concise, structured narrative text summary (≤ 30 lines) optimized for insertion directly into the LLM context window.
  • Structured Content Channel (structuredContent): A complete JSON envelope containing raw metric arrays, data provenance, freshness timestamps, and quality alerts, rendered visually by MCP App widgets in compatible host interfaces.

​
Next Steps & Cross-References

Semantic Layer

Review canonical marts and ratio calculations supporting agent queries.

Anomaly Detection

Understand z-score thresholds and severity levels for anomaly alerts.

Security Constraints

Examine FastMCP dual-channel response rules and tenant isolation.

Quickstart Guide

Test MCP tool responses locally with the FastMCP dev server.