How does a wrong answer become a test?
Open a trace, the record of one request and its steps. Check what the app found and what it answered. When you create a dataset example from the trace, Opik includes the input and stores the actual output as the expected output. Replace a wrong output with the result you actually want.
Does a masked display remove the original data?
Opik documents Enterprise preview masking that can hide selected data on screen. All default workspace roles, including Read, can see the original data. To mask it for someone, an administrator must give that person a custom role without permission to see originals. The original trace remains stored, and attachments are among the exceptions. Leave private fields out of the trace when they do not need to be there.
How can one request count as several steps?
A single request can include separate recorded steps for search and generation. Comet lists span limits and retention for Opik Cloud, and a self-hosted option where you operate the storage. Save the test cases and scoring rules. Opik documents project exports for datasets, experiments and traces.
What should the saved test expect?
Draw a fictional booking assistant whose trace shows the correct closing time but whose answer gives the wrong time.
- Find the first wrong stepMark the source lookup correct and the final reply wrong.
- Make the example repeatableWrite the expected closing time from the fictional source. Keep it separate from the saved wrong output.