EXPERIMENTAL recall preflight for physical posting search — massive_posting_recall

Checks evenly spaced, distant-from-center, and small-posting source rows against streamed exact neighbors before writing a full graph. At least one audit row represents a small nonempty posting list; larger audits reserve up to one quarter for these lists. Center scoring uses bounded candidates when a full scan is too costly. The sampled recall cannot certify unsampled rows.

Usage

massive_posting_recall_pilot(
  x,
  postings,
  centers,
  k = 30L,
  nprobe = NULL,
  sample_rows = 32L,
  target_recall = 0.99,
  reference_chunk_rows = 8192L,
  n.cores = 1L,
  memory_limit = "1GB"
)

Arguments

x: Original file-backed matrix used to build postings.
postings: A result of massive_coarse_postings().
centers: Centers used to build postings.
k: Number of non-self neighbors.
nprobe: Number of posting lists searched per sampled row.
sample_rows: Number of sampled rows, at most 256.
target_recall: Minimum required recall for each sampled row.
reference_chunk_rows: Maximum reference rows read at once.
n.cores: CPU workers for distance calculations.
memory_limit: Conservative algorithm-buffer RAM budget.

Value

Sampled row IDs and recalls, summary, backend, and estimated exact-reference read and comparison counts.