Evaluation results
Enter ranked predictions and relevant items, then calculate Recall@K.
Recall across K
Found versus missed
Per-query recall at largest K
Relevant-item breakdown
Relevant items found
Relevant items missed
Detailed results
No rows| Query | Group | Model | K | Top-K size | Found | Total relevant | Missed | Recall |
|---|---|---|---|---|---|---|---|---|
| Results will appear here. | ||||||||
Calculation history
Recent summaries are stored in this browser.
Formula used
Recall@K = Relevant items retrieved in the top K ÷ Total relevant items
The value ranges from zero to one. Multiplying by 100 produces a percentage.
How to use this calculator
- Choose single-query or batch evaluation mode.
- Enter ranked predictions and ground-truth relevant items.
- Add one or more positive K values.
- Adjust matching, duplicate, tie, and averaging behaviour.
- Calculate, inspect charts, and export the detailed results.
Example data
| Input | Example |
|---|---|
| Ranked predictions | A, B, C, D, E |
| Relevant items | B, D, F |
| K | 3 |
| Relevant found | B |
| Recall@3 | 1 ÷ 3 = 0.3333 = 33.33% |
Metric guidance
Recall@K rewards systems that retrieve more relevant items early. It does not penalise irrelevant items directly. Pair it with Precision@K for balance.
Hit Rate@K only checks whether one relevant item appears. Recall@K measures the share of all relevant items recovered. Average Precision also considers relevant-item positions.
Choose K from the number of items users can realistically inspect. Compare several K values before selecting one. Report both aggregate and per-query results.
Frequently asked questions
What does Recall@K measure?
It measures the proportion of all relevant items retrieved within the first K ranked results.
Is a higher Recall@K better?
Yes. A higher value means more ground-truth relevant items appeared before the selected cutoff.
Can Recall@K exceed one?
No. Properly calculated recall remains between zero and one, inclusive.
How is Recall@K different from Precision@K?
Recall focuses on recovered relevant items. Precision focuses on how many top-K results are relevant.
What happens when K exceeds the ranking length?
The calculator uses all available ranked items and displays a warning.
How are duplicate predictions handled?
You may keep the first occurrence, keep the last occurrence, or preserve ranking positions.
What is graded relevance?
Graded relevance assigns numeric importance scores. The threshold decides which items count as relevant.
What is macro-average recall?
Macro recall averages query-level recall values equally, regardless of relevant-set size.
What is micro-average recall?
Micro recall divides total relevant items found by total relevant items across queries.
When should weighted recall be used?
Use it when some queries, users, or business segments deserve greater influence.