# LABELED FOLLOW-UP: C5 k_bits spot-check. The record run capped SIFT k_bits
# at 256 (dim-scaled: {64,128,256} for 128d) and
# recall proved k_bits-BOUND, not probe-bound (scan_fraction 0.02->0.126
# moved recall 0.751->0.742 while p50 went 0.40->1.00ms) -- k256's ceiling
# was 0.925 @ 1.46ms vs the 0.95 @ <=1.2ms bar. This tests k512@128d (the
# k_bits-dominates finding says wider codes are latency-free until Hamming
# surfaces above the fetch cost) plus a low-fraction axis at the fixed
# low pool, since C5's budget needs BOTH knobs at once.
#
# Main dataset section = the (cached) hard corpus with no systems: the
# runner requires a dataset, SIFT runs via the sift1m section.
name: followup_c5_sift_k512
seed: 42
k: 10
warmup_queries: 20
hardness_recall_threshold: 0.97
min_throughput_seconds: 5.0

postgres:
  maintenance_work_mem: 8GB
  min_wal_size: 1GB
  max_wal_size: 4GB

dataset:
  kind: synthetic
  n: 1000000
  dim: 768
  n_centers: 1000
  cluster_sigma: 1.1
  cluster_scale_mix: [0.6, 1.0, 1.6]
  background_frac: 0.02
  n_queries: 500

metric: l2

filters:
  selectivities: []
  modes: []
  run: []

systems: []

concurrency: [1]

churn:
  enabled: false

sift1m:
  enabled: true
  sift_n_queries: 500
  sift_systems:
    - name: pnat_sift_k512
      type: pg_pinecone_native
      build_params: {seed: 42, k_bits: 512}
      sweep_param: rescore_multiplier
      sweep_values: [1, 2, 3, 4, 10]
    - name: pnat_sift_k512_frac
      type: pg_pinecone_native
      build_params: {seed: 42, k_bits: 512, rescore_multiplier: 2}
      sweep_param: scan_fraction
      sweep_values: [0.02, 0.04, 0.08]
