Recipe: score a structured-output prompt
Recipe — score structured outputs. Pointer to the SDK document-evaluation sample.
Last updated
Was this helpful?
Recipe — score structured outputs. Pointer to the SDK document-evaluation sample.
For structured-output evaluation (JSON, fields, schemas), the canonical samples are:
samples/modalities/document_evaluation.py — extraction accuracy, cross-field consistency, structural integrity
samples/core/judge_creation_and_test.py — build a custom judge that validates structured output against rules
Define a judge with client.judges.create(name=, evaluation_goal=) whose rubric encodes your schema requirements. Run on traces via client.trace_evaluations.create(trace_id=, judge_id=).
For deterministic JSON-schema validation specifically, validate client-side before posting traces — that catches shape errors before consuming an evaluation.
Last updated
Was this helpful?
Was this helpful?