Requests over time
Every question, grouped by what happened to it
Where answers came from
Fresh model answers versus cache hits, refusals and failures
Response time
Average and slowest 5% (p95) per period, in seconds
Token usage over time
Input tokens billed in full, input served from OpenAI's prompt cache (about half price), and output
Where the time goes
Average seconds per stage, for freshly generated answers
Tokens by stage
Router, answer model, cache embedding and cache verifier
Response-time spread
How many requests finished in each time band
What people ask about
Topics the router matched (customer, repair, fault, cost...)
Models and cost
Estimated from your price settings in config.json
Failures and issues
Requests that failed, and SQL the safety check rejected or the database refused
Request log
| When | Question | Outcome | Time | Tokens in | out | Cost | SQL |
|---|