EXPERIMENTAL measured recall for file-backed landmark neighbors — massive_audit_landmark

Compares evenly spaced query rows with an independent streamed exact CPU search. Sampled queries share bounded reference scans; the number of passes is reported in audit_reference_passes. This audit can still be expensive and does not certify unsampled rows or change the saved neighbor files.

Usage

massive_audit_landmark_knn(
  knn,
  query,
  reference,
  sample_rows = 32L,
  memory_limit = "1GB"
)

Arguments

knn: A result of massive_landmark_knn().
query, reference: Original file-backed matrix descriptors.
sample_rows: Number of query rows to audit (at most 10,000).
memory_limit: RAM budget for exact-search buffers.

Value

knn with sampled mean and minimum row recall, a count of sampled rows below the target, and sampled row IDs in metadata.