Configure
Target, filter, sampling, return type, and passing range for an evaluator.
Every evaluator shares the same configuration, plus a few fields specific to its type.
Shared settings
| Setting | What it does |
|---|---|
| Target | Score the whole run or an individual span. |
| Event filter | Only score runs or spans that match these conditions. |
| Sampling | Score a percentage of matching events (1 to 100). |
| Return type | numeric, boolean, or string (category). |
| Passing range | Defines passed: a numeric range, the passing boolean value, or the set of passing strings. |
| Enabled | Score live ingestion now, or keep as a draft. |
Type-specific config
- LLM judge: provider and model, the model integration (your key), system prompt, user template, and output parser.
- Code: the source code and runtime version.
- Human: grading instructions (the UI mode is derived from the return type).
- Composite: the source evaluators, the aggregation, and any weights.
Versioning
Editing the rubric (the prompt, code, return type, or passing range) bumps the evaluator's version, so prior results stay valid against the rubric that produced them. Toggling enabled on or off does not bump the version.