Analytics & Insights

Two related but distinct capabilities: usage analytics (what's being queried, by whom, how fast) and the AI insights engine (what the data itself is telling you).

Usage Analytics

Requires analytics.enabled: true. Exposed via /api/analytics/*:

Endpoint	Reports
/summary	Total queries, success rate, overview stats
/hourly-stats	Query volume by hour
/errors	Failed queries and error patterns
/widget-comparison	Standard widget vs. auth widget usage
/performance	Query execution time distribution
/top-users	Highest-volume users/sessions
/query-complexity	JOIN count, subquery depth distribution

This tracks usage patterns. For per-call token cost and risk scoring, see Audit Logs instead — the two are separate systems (analytics is aggregate/optional, audit logging runs for every LLM call regardless of the analytics flag).

AI Insights Engine

Runs after query execution to surface patterns a user might not think to ask about:

insights:
  enabled: true
  mode: hybrid          # llm_only | statistical_only | hybrid
  max_insights: 3
  include_statistical: true

Mode	Behavior
llm_only	Insight generation delegated entirely to the LLM (uses the insights_generation prompt)
statistical_only	Pure statistical analysis (trend/anomaly detection), no LLM call — cheaper, faster
hybrid	Statistical signals feed into the LLM prompt for richer, still-grounded insights

Insights are context-aware: they account for temporal patterns (daily/weekly/monthly cycles) and flag when a query touches incomplete real-time data (e.g. "today" partition still filling in) rather than misreading a partial day as a trend.

Each insight carries a type (trend / anomaly / recommendation / summary) and severity (low / medium / high), and the prompt driving generation is fully editable — see Runtime Prompts.