# Forced-seqscan exact baselines on the REAL 768d corpus (the real-corpus
# companion of baselines_exact.yaml). Same series construction as
# configs/baselines_exact.yaml (see that file's header for the methodology);
# only the dataset differs.
#
# CORPUS: facebook/wiki_dpr -- DPR-encoded English Wikipedia passages, 768d
# fp32, REAL embeddings at the headline dimensionality (no dimension
# deviation). Publicly downloadable without auth from
# dl.fbaipublicfiles.com/dpr/data/wiki_encoded/single/nq/wiki_passages_{0..3}
# (4 shards, 5.3GB total, within the 6GB budget; the Cohere wikipedia-22-12
# sets are gated on HF as of 2026-09, and CohereLabs/wikipedia-2023-11 is
# 1024d). Shards were converted once to datasets_cache/wikidpr_768d_1m.npz:
# first 1,000,000 passages in corpus order as the base set; from the
# REMAINING 681,224 passages a seed-42 shuffle takes 500 measured queries,
# 160 warmup vectors, and a 100,000-vector query bank -- all held out, all
# real. Converter provenance (shard URLs, content sha) is embedded in the
# artifact's meta_json and stamped into the dataset key.
#
# Geometry sanity (measured at conversion): norms mean 12.22 std 0.36
# (min 10.98, max 14.98 -- DPR vectors are unnormalized, trained for inner
# product); pairwise distance over a 500-point sample: mean 12.44, p5 10.80,
# p95 13.96. Metric here is l2 for consistency with every other corpus here; the corpus geometry (not the training objective) is what the
# baseline measures.
#
# FILTERS: cluster structure for the filter matrix comes from k-means
# (k=1000, numpy Lloyd on a 100k sample, 15 iters, seed 42, cached beside the
# artifact): correlated = whole k-means clusters until target selectivity
# (same correlated_chosen_clusters rule as synthetic), uncorrelated = random.
# In-region ("tenant") queries are REAL held-out bank vectors whose nearest
# centroid is inside the qualifying region. NOTE: the 100k bank holds ~1000
# in-region vectors at 1% selectivity (enough for the 660-query pool) but
# only ~100 at 0.1% -- there the pool is sampled WITH REPLACEMENT and the
# results JSON records n_unique_available per filter (in_region_sampling).
#
# No churn (real corpora have no generative model to re-draw from -- the
# harness hard-errors if enabled), no SIFT (independent sanity anchor,
# already run).
name: baselines_exact_real
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: real
  artifact: wikidpr_768d_1m.npz
  n: 1000000
  dim: 768
  n_centers: 1000
  kmeans_sample: 100000
  kmeans_iters: 15
  n_queries: 500

metric: l2

filters:
  selectivities: [0.001, 0.01, 0.10, 0.50]
  modes: [correlated, uncorrelated]
  run:
    - {selectivity: 0.001, mode: correlated}
    - {selectivity: 0.001, mode: uncorrelated}
    - {selectivity: 0.01, mode: correlated}
    - {selectivity: 0.01, mode: uncorrelated}
    - {selectivity: 0.10, mode: correlated}
    - {selectivity: 0.10, mode: uncorrelated}
    - {selectivity: 0.50, mode: correlated}
    - {selectivity: 0.50, mode: uncorrelated}

systems:
  - name: exact_seqscan
    type: exact
    sweep_param: none
    sweep_values: [0]

concurrency: [1, 8]
