Overview
Natural-language questions translated to SQL against your telemetry, with clarifying responses for ambiguous input rather than silent empty results.
A question goes through a prompt carrying your telemetry schema, generation, reflection retries on malformed output, then read-only execution and a summarisation pass. Follow-ups are rewritten to be self-contained, so "what type of errors are these?" resolves against the previous turn.
Key features
- Natural language → SQL, shown to you before it runs
- Read-only execution — multi-statement support disabled
- Ambiguous questions get a clarifying response, never a best guess
- Multi-turn context rewriting for follow-up questions
Example
Question · Which services have the slowest p95 latency in the last hour?
SELECT service_name,
APPROX_PERCENTILE(duration / 1000000, 0.95) AS p95_ms,
COUNT(*) AS spans
FROM otel_traces
WHERE timestamp >= NOW() - INTERVAL 1 HOUR
GROUP BY service_name
ORDER BY p95_ms DESC
LIMIT 5;| service_name | p95_ms | spans |
|---|---|---|
| checkout | 842.6 | 18,902 |
| payment | 611.3 | 12,440 |
| product-catalog | 388.9 | 96,715 |
| cart | 124.2 | 44,108 |
| ad | 98.7 | 9,336 |
Answer · checkout is the slowest at 842 ms p95, roughly 1.4× payment. Both sit on the same downstream call to product-catalog, which is itself elevated.
Delivery
Ask AI is part of the full platform in both delivery models — there is no cut-down edition.
SaaS · Fully managed
We run the platform, handle upgrades, scaling and backups. You point your collectors at an endpoint and start asking questions the same afternoon.
PaaS · Your infrastructure, our platform
The full platform deployed into your own Kubernetes clusters or VMs, with telemetry never leaving your network boundary. We operate and support it where you host it.