Module ann_benchmark 

ANN (Approximate Nearest Neighbour) recall and latency benchmarking.

Provides utilities for evaluating the quality and performance of ANN indices: recall@k, QPS, build time, memory usage, ground-truth generation, precision-recall tradeoff, latency percentiles, and report generation.

Structs

BenchmarkReport

A complete benchmark report.

BuildTimeResult

Result of a build-time measurement.

BuildTimer

Track how long an index build takes.

PrecisionRecallPoint

A data point on the precision-recall curve.

QpsResult

Result of a QPS measurement.

Functions

average_distance_ratio

Compute the average distance ratio: sum of approx_dist / true_dist for each query’s k-th neighbour. A perfect index has ratio = 1.0.

brute_force_knn

Brute-force computation of exact k nearest neighbours for a set of queries.

estimate_flat_memory

Estimate the memory footprint of a flat vector index (vectors only).

estimate_hnsw_memory

Estimate memory for an HNSW-like graph index.

estimate_pq_memory

Estimate memory for a product-quantised (PQ) index.

measure_qps

Measure queries per second for a given search function.

per_query_recall

Compute recall@k for individual queries (not averaged).

precision

Compute precision: the fraction of returned results that are true neighbours.

precision_recall_sweep

Run a sweep over a set of parameter values and collect recall/precision at each setting.

recall_at_k

Compute recall@k: the fraction of true k-NN that appear in the approximate result set.

Type Aliases

Neighbour

A single query result: (vector_id, distance).