Sample readiness
The sample target is met. Automatic selection or a manual run can start the evaluation.
These checks apply to every experiment for knowledge-agent, whatever the source model. Changes apply to new runs and start a fresh evidence pool, so earlier runs are not mixed with runs under different rules.
Added to the judge's instructions. A candidate that violates them is judged not equivalent.
Checked without the judge. A candidate that fails it counts as not equivalent and skips the judge call. If the original output fails it too, the schema is likely wrong, so the judge decides and the miss is only reported.
Test whether knowledge-agent answers as well with a shorter prompt: less conversation history, or a shorter system prompt that the judge model proposes. TrAIce replays 30 captured requests on gpt-5.5 with the shorter prompt, and the judge compares each answer with the one your application got. Nothing changes in your traffic.
This demo is read-only. Create a workspace to run a trial.
Use the proposed system prompt
Across 2 trials of the proposed system prompt, 59 of 60 answers were as useful as the full prompt's (91% to 100% at 95%). Requests sent 7% fewer input tokens. At this workload's uncached input over the last 30 days, that saves about $541.45 a month.
Opening the prompts is recorded in the audit log. Kept until Nov 20, 2026.
Once a week, TrAIce groups up to 500 of this workload's sampled requests by what they ask for and shows what each group costs. The judge model on Google Vertex AI names each group from a few of its requests. Names are kept only as long as the samples they came from (at most 14 days); the counts and costs stay.
No analysis yet. The first one runs within a day or so.
These complete, unexpired prompt and response pairs can be selected for an experiment. Opening sample content is audited.
| Case | Captured | Prompt tokens | Response tokens | Production latency | Expires | Availability | Sample details |
|---|---|---|---|---|---|---|---|
| 31 | Nov 8, 08:12 PM |
| Started | Change | As useful | 95% interval | Result | Input tokens | Saving per month | Cost | Cases |
|---|---|---|---|---|---|---|---|---|
| Nov 6, 2026 | Proposed system prompt | 29 of 30 | 83% to 100% | Completed | 7% fewer | About $541.45 | $0.74 | |
| Nov 2, 2026 | Proposed system prompt | 30 of 30 | 88% to 100% | Completed | 7% fewer | About $546.69 | $0.88 |
| 72339 |
| 3049 |
| 1,213 ms |
| Nov 22, 08:58 PM |
| Fresh for next run |
| View sample |
| 32 | Nov 8, 08:11 PM | 163550 | 8944 | 2,268 ms | Nov 22, 08:58 PM | Fresh for next run | View sample |
| 33 | Nov 8, 08:10 PM | 353833 | 15637 | 3,693 ms | Nov 22, 08:58 PM | Fresh for next run | View sample |
| 34 | Nov 8, 08:09 PM | 125807 | 4503 | 1,464 ms | Nov 22, 08:58 PM | Fresh for next run | View sample |
| 35 | Nov 8, 08:08 PM | 41516 | 1923 | 598 ms | Nov 22, 08:58 PM | Fresh for next run | Close |
Sample case 35 evidenceCaptured prompt and source response available for the next experiment. fresh Captured promptNorthstar support case 35: resolve the knowledge-agent request with the account policy and cited help-center context. Original responsegpt-5.5 Resolved knowledge-agent case 35 with the correct policy, next step, and customer-safe explanation. | |||||||
| 36 | Nov 8, 08:05 PM | 78630 | 3362 | 406 ms | Nov 22, 08:58 PM | Fresh for next run | View sample |