Full Recall¶
Measures whether all relevant document groups were retrieved (all-or-nothing).
Overview¶
| Field | Value |
|---|---|
| Type | Retrieval |
| Range | {0, 1} |
| Higher is better | Yes |
Formula¶
\[FullRecall@k = \begin{cases} 1.0 & \text{if every GT group has at least one hit in Retrieved@k} \\ 0.0 & \text{otherwise} \end{cases}\]
The final score is the average of Full Recall across all queries.
Interpretation¶
| Score | Meaning |
|---|---|
| 1.0 | All evidence groups retrieved |
| 0.0 | At least one evidence group missing |
Unlike standard Recall which measures the fraction of ground truth groups found, Full Recall is a strict binary measure -- a query scores 1.0 only when every group is satisfied.
When to Use¶
Use when complete evidence coverage is required, such as:
- Multi-hop question answering (all supporting facts must be retrieved)
- Fact verification requiring multiple evidence pieces
- Measuring the "success rate" of queries with full coverage
Configuration¶
_target_: autorag_research.evaluation.metrics.retrieval.FullRecallConfig