Per-Embed Metrics & Analytics Dashboard
Granular performance telemetry for every deployed assistant: impressions, opens, chats, messages, containment, unanswered queries, fallback reasons, and lead conversion.
Session & Funnel Metrics
- Impressions → opens → chats → messages funnel
- Containment & refusal distribution
- Lead prompt performance & abandonment
- Time-to-first-answer latency histograms
Retrieval & Quality Signals
- Fallback reason codes (low score, timeout, provider error)
- Unanswered / escalation intent surfacing
- Top cited documents & stale content alerts
- Precision / hallucination sampling workflow
Operational Insight
Per‑embed isolation allows product managers and agencies to compare performance across properties or clients. Export raw events for deeper cohort or attribution modelling.
All metrics are traceable—each aggregated datapoint ties back to raw events (OpenTelemetry spans for retrieval + generation) ensuring auditability and regression investigation.
FAQ
What is a containment metric?
Containment measures sessions resolved without human escalation; it reflects deflection efficiency and assistant trustworthiness.
How often are analytics updated?
Event ingestion is near real-time (<5s typical) with aggregation windows rolled up in minute/hour/day buckets.
Can we export raw events?
Yes—CSV / NDJSON export plus roadmap for streaming export to your data warehouse.