A retrieval system can fail quietly when it supplies partial evidence: models answer incorrectly more often with insufficient context than with no context. section 14.5.2
Sufficient context means that the retrieved set collectively contains the information needed to answer the query. It differs from relevance: relevant passages may still fail to justify the answer. section 14.5.2
Use the metric to guide evaluation and response policy:
- Compare retrieval configurations rather than treating an absolute score threshold as universal. section 14.5.2
- Use sufficiency judgments to decide when the system should abstain. section 14.5.2
- Check corpus state separately: a superseded policy can answer the question completely while remaining outdated. section 14.5.2
Why partial evidence is dangerous
Gemma answered incorrectly on 10.2% of questions without context and 66.1% with insufficient context. A prompted autorater classified sufficiency correctly at least 93% of the time; abstention based on its labels raised answer correctness by 2–10 points at the same coverage. section 14.5.2
