recall at k

**Recall at k** is the **retrieval metric that measures whether relevant documents are present within the top-k returned results** - it quantifies coverage of needed evidence. **What Is Recall at k?** - **Definition**: Proportion of relevant items recovered in the first k retrieved candidates. - **Binary Variant**: For single-answer tasks, often treated as hit or miss at top-k. - **Sensitivity Profile**: Emphasizes not missing relevant evidence, regardless of rank position within k. - **RAG Relevance**: High recall is prerequisite for answerable grounded generation. **Why Recall at k Matters** - **Answer Feasibility**: If no relevant passage is retrieved, generation cannot be reliably correct. - **Retriever Coverage**: Detects blind spots in query understanding and index representation. - **Model Comparison**: Useful first-pass metric for candidate retriever evaluation. - **Pipeline Tuning**: Guides top-k size and hybrid retrieval design choices. - **Safety Role**: Better recall reduces unsupported fallback to parametric guesses. **How It Is Used in Practice** - **k Sweep Analysis**: Measure recall across multiple k values to find diminishing returns. - **Segment Diagnostics**: Break down recall by query type and domain difficulty. - **Joint Evaluation**: Pair with precision and rank metrics for balanced optimization. Recall at k is **a foundational coverage metric in retrieval evaluation** - strong recall is essential to ensure relevant evidence is available for downstream grounded answer generation.