Recall@K

Recall@K measures coverage—the fraction of all relevant items successfully retrieved in the top K results.

Formula: Recall@K = |{Relevant Items} ∩ {Top K Items}| / |{Relevant Items}|

• Numerator: Number of relevant items that appear in the top K positions
• Denominator: Total number of relevant items in the ground truth
• Range: 0 to 1 (or 0% to 100%)
• A score of 1.0 means all relevant items were surfaced within the top K

Recall@K and Precision@K are complementary metrics that trade off against each other.

• Recall@K: "Did we find all the good stuff?" — prioritizes completeness
• Precision@K: "How much of what we showed was good?" — prioritizes accuracy

A system returning 100 items to guarantee capturing all 5 relevant items has high recall but low precision. Conversely, a conservative system returning only 2 highly-confident items may have high precision but low recall.

The choice of K dramatically impacts interpretation and should align with the user interface constraints.

• K=5 or K=10: Common for top-N recommendation lists where screen real estate is limited
• K=20 or K=50: Used in candidate retrieval stages before re-ranking
• K=100+: Relevant for search engines measuring first-page recall

Rule of thumb: Match K to the number of items a user can realistically consume in a single viewport or interaction.

Recall@K treats all positions within K as equally valuable, unlike Normalized Discounted Cumulative Gain (NDCG) which applies a logarithmic position discount.

• Recall@K: Binary relevance within top K — position 1 and position K are weighted identically
• NDCG@K: Position-sensitive — a relevant item at rank 1 contributes more than one at rank K

When to prefer Recall@K: When the primary goal is ensuring no relevant item is missed, regardless of ordering within the retrieved set.

Recall@K is the primary metric for candidate generation in two-stage recommender architectures.

Stage 1 (Candidate Retrieval):

• Goal: Maximize Recall@K (typically K=100 to K=1000)
• Model: Two-tower networks, ANN indices like HNSW
• Tolerance: High recall is critical; precision is secondary

Stage 2 (Ranking):

• Goal: Optimize NDCG or CTR
• Model: Deep crossing networks, attention mechanisms
• The ranker can only work with what the retriever surfaces — making Recall@K a hard upper bound on final system performance.

Recall@K has several known limitations that practitioners must account for:

• Ignores ranking quality: A relevant item at position 1 and position K are treated identically
• Requires complete relevance judgments: Missing annotations inflate the denominator, artificially lowering recall
• Doesn't capture user satisfaction: A list with perfect recall may still deliver poor user experience if top positions are irrelevant
• Scale sensitivity: As the total number of relevant items grows, achieving high recall becomes harder

Mitigation: Pair Recall@K with position-aware metrics like MRR or NDCG for a complete evaluation picture.

A ranking quality metric that measures the usefulness of a ranked list, giving higher weight to relevant items appearing at the top positions. Unlike Recall@K, NDCG is position-sensitive—a relevant item at rank 1 contributes more to the score than one at rank 10. The metric normalizes against an ideal ranking, producing a value between 0 and 1. Commonly used when relevance has graded levels (e.g., 'highly relevant' vs 'somewhat relevant') rather than binary labels.

The complementary metric to Recall@K that measures the proportion of recommended items within the top K that are actually relevant. While Recall@K answers 'Did we find all the good stuff?', Precision@K answers 'How much of what we showed was actually good?' The two metrics exhibit an inherent trade-off—increasing K typically raises recall but lowers precision. Used together, they provide a complete picture of retrieval quality.

A class of algorithms that trade a small amount of accuracy for massive speedups in finding the closest vectors to a query embedding. Essential for serving Recall@K in production at billion-item scale. Key implementations include:

• HNSW: Multi-layered proximity graphs for logarithmic search complexity
• FAISS: Facebook's GPU-optimized library for billion-scale similarity search
• ScaNN: Google's ANN with anisotropic vector quantization Without ANN, exact nearest neighbor search would be computationally prohibitive for real-time retrieval.

A training efficiency technique that approximates the full softmax over a massive item catalog by updating only the weights for the positive target item and a small, randomly selected set of negative examples during each training step. This directly impacts Recall@K performance—the choice of negative sampling strategy (uniform, in-batch, hard negatives) significantly influences whether the model learns to distinguish truly relevant items from superficially similar distractors.

An evaluation metric that measures how high the first relevant item appears in the ranked list. Calculated as the average of the reciprocal of the rank position of the first correct answer across all queries. MRR is particularly useful when users typically need only one relevant result (e.g., 'find the exact product I want'), whereas Recall@K is preferred when surfacing all possible candidates matters (e.g., 'show me all compatible accessories').

A binary metric that measures whether at least one relevant item appears anywhere in the top K recommendations. Less granular than Recall@K—it doesn't distinguish between finding one relevant item or finding all of them. Often used as a simpler, more interpretable complement to Recall@K in A/B testing reports. A model with high Hit Rate@10 but low Recall@10 is finding something relevant but missing most of the catalog.