Embedding-Eval Palace
8 cardsFront foyer
Recall@k
confirmedA fisher casting a net of exactly k holes at the foyer door — Recall@k is how many of the house's fish the net actually caught.
Show definition
The fraction of all relevant documents that appear in the top-k retrieved results. It answers: of everything I should have found, how much did the top k surface?
Precision@k
confirmedA picky doorman at the foyer inspecting exactly k guests — Precision@k is how many of those k he correctly waved through.
Show definition
The fraction of the top-k retrieved results that are actually relevant. It answers: of what I showed the user in the top k, how much was on target?
Long hallway
MRR
confirmedIn the long hallway, doors are numbered 1, 2, 3. MRR shouts down the corridor and scores you by one-over the number of the FIRST door that opens onto a right answer.
Show definition
Mean Reciprocal Rank: the average of 1 divided by the rank of the first relevant result across all queries. It rewards putting a correct answer near the very top.
NDCG
needs reviewA grand staircase in the hallway where each higher step is worth logarithmically less — NDCG sums the gems on each step, then divides by the treasure of a perfectly stacked staircase.
Show definition
Normalized Discounted Cumulative Gain: a graded-relevance ranking metric that discounts each hit by the log of its position, then normalizes against the ideal ordering so the best possible ranking scores 1.0.
Sunny kitchen
MAP
confirmedA treasure MAP pinned in the kitchen: at each spot you dig up a relevant find you note the precision so far, then average all the notes — and average across every map you own.
Show definition
Mean Average Precision: for each query, average the precision measured at every rank where a relevant document appears, then average those per-query scores across all queries.
Hit Rate
confirmedA kitchen dartboard: Hit Rate only cares whether at least one dart of the k you threw landed anywhere on the board — a plain hit-or-miss tally across all your throws.
Show definition
The fraction of queries for which at least one relevant document appears in the top-k results. A coarse, binary 'did we get anything right?' signal per query.
Book-lined study
Cosine Similarity
suspendedTwo reading lamps in the study angled toward each other — cosine similarity is how aligned their beams are, caring about angle, never about brightness.
Show definition
The cosine of the angle between two embedding vectors, equal to their dot product divided by the product of their magnitudes. It scores directional closeness independent of vector length.
Attic bedroom
P95 Latency
confirmedAn attic bedroom hourglass where 95 of every 100 sand grains have already fallen — P95 latency is the mark the slowest tail-enders still beat.
Show definition
The 95th-percentile response time: the latency value below which 95% of requests complete. It captures tail slowness that an average would hide.
Your learning and review progress is stored only on this device.