Regressa

Evals

Eval definitions live per project and optionally scope to one prompt template. sample_rate controls the fraction of matching traces that get scored.

type config how it scores
llm_judge { rubric, judge_model?, pass_threshold } Judge model (default claude-sonnet-5) scores 0-100 via a forced tool call; score/100
semantic_similarity { pass_threshold, embedding_model } + golden examples Best cosine similarity between output embedding and golden outputs
custom_function { source, pass_threshold } JS function body run in node:vm with a 250ms timeout; receives trace

Results land in eval_results with score (0-1), passed, reasoning, and the eval's own cost.