You are building an automatic eval for a RAG system’s l...

Grading long-form answers as a single unit is ineffective because it collapses partial successes and failures into a binary result. The correct approach is to decompose the response into atomic facts, verify each fact individually, and then aggregate the results. This provides a granular view of performance rather than a blunt, all-or-nothing score.