Assistant metrics and SQL
Define metrics precisely and query them consistently.
Data and BIProductQA
Definitions
| Metric | Definition |
|---|---|
| Containment | Share not escalated to a person |
| Resolution | Goal met, by confirmation or no repeat contact |
| Repeat contact | Same user and topic within 7 days |
| Fallback rate | Share of turns hitting not-understood or not-found |
| Groundedness | Share of sampled answers supported by sources |
| Latency | p50 and p95, not averages |
| Cost per resolved | All costs, including humans, divided by resolved conversations |
Log at minimum
- conversation_id, turn_id, hashed user id, timestamp
- channel, language, market
- route, confidence, topic
- model, prompt and knowledge-base versions
- retrieval ids and scores; tool calls and status
- latency parts, tokens, costs
- outcome, escalation reason, feedback, guardrail triggers
Cost per resolved (SQL)
SELECT route, SUM(model_cost + tool_cost + retrieval_cost + human_minutes * 0.75) / NULLIF(SUM(CASE WHEN resolved THEN 1 ELSE 0 END), 0) AS cost_per_resolved FROM conversations GROUP BY route;
Dashboard rules
- Segment by language, channel, topic and complexity
- Show distributions and sample sizes
- Mark release dates
- Every chart links to an owner and an action