A measure of approximate (ANN) vector search accuracy: what fraction of the true nearest neighbours, as found by an exact kNN search, the approximate search returned. A value of 1 means nothing was approximated.
Read more: Microsoft Learn
In the Ultra Transcenders books
Each book explains Recall in context, with comparison tables and the common traps.
Terms in this definition
- Measure
A figure like total revenue, worked out by aggregating fact table data wherever dimensions intersect within a semantic model.
- ANN
Short for approximate nearest neighbour: a vector search that finds most of the closest matches rather than guaranteeing every one, trading completeness for speed and lower cost. DiskANN and HNSW indexes work this way, whereas flat search is exact.
- Vector search
Searching embeddings for similar meaning, so paraphrases are found via nearest-neighbour algorithms like HNSW; exact identifiers, however, may be missed.
- Exact nearest neighbour
Measures how far the query vector is from each and every candidate, for instance using VECTOR_DISTANCE with TOP and ORDER BY, so the results are always right. It is advised for up to roughly 50,000 candidates; beyond that, ANN (approximate) search is preferred.
Related terms
- HNSW
Graph-structured index for approximate vector search in RediSearch and pgvector. Compared with IVFFlat, recall versus speed is better, though building takes longer and memory use is higher.
- IVFFlat
An approximate nearest-neighbour index in pgvector that groups vectors into lists. Compared with HNSW it builds faster and takes less space but balances speed and recall less well, and data must be loaded before it is created.
- Response Completeness evaluator
Evaluator in preview for RAG scenarios: it asks whether the answer contains the essential facts from the ground truth, so it measures recall, complementing groundedness, which measures precision.