Combined evaluation scores several quality dimensions in one judge prompt or evaluation call. It improves efficiency when dimensions are relatively independent, but requires prompts that clearly separate and independently assess each dimension. section 4.4.3
Design a combined evaluation by following these checks:
- Name each dimension separately, such as helpfulness, accuracy, tone, and policy compliance. section 4.4.3
- Instruct the judge to assess dimensions independently; generating reasoning steps for each dimension can help maintain separation. section 4.4.3
- Prefer sequential evaluation when dimensions interact complexly or need specialized prompts; prefer combined evaluation when minimizing cost matters. section 4.4.3
- Weight dimensions by their impact instead of averaging them naively; use a geometric mean when principled weights are unavailable and severe failures should dominate. section 4.4.3

