# Unfiltered local engines at wiki_dpr 1M x 768d: pgvector hnsw and ivfflat, pg_pinecone native, forced exact seqscan.
# recall@10 and recall@100 against exact ground truth, latency p50/p95/p99 and QPS at 1, 4 and 16 clients, build time, index
# size. Same artifact, seed, k-means and ground-truth caches as configs/residual-c1.yaml (dataset key real_fd937e5862449934).
# configs/full_1m.yaml is the 2026-09 synthetic run (hnsw and
# ivfflat only, churn, SIFT, 6.1 h) and measures neither native, exact nor wiki_dpr.
# warmup_queries is 10 because 16 clients x 10 = the 160 warm-up vectors the artifact holds.
# The knob grids follow wikidpr_1m_pg_pinecone.yaml; ef_search must be >= k to return k rows,
# so k = 100 sweeps its own values (sweep_values_by_k).
name: local_1m
seed: 42
k: 10
k_values: [10, 100]
warmup_queries: 10
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:
  run: []

systems:
  - name: pgvector_hnsw
    type: hnsw
    build_params: {m: 16, ef_construction: 64}
    sweep_param: ef_search
    sweep_values: [10, 20, 40, 80, 160, 320]
    sweep_values_by_k: {100: [100, 160, 320, 640]}

  - name: pgvector_ivfflat
    type: ivfflat
    build_params: {lists: null}
    sweep_param: probes
    sweep_values: [1, 2, 4, 8, 16, 32, 64]

  - name: pg_pinecone_native
    type: pg_pinecone_native
    build_params: {seed: 42}
    sweep_param: rescore_multiplier
    sweep_values: [1, 2, 4, 10, 20, 50]

  - name: exact_seqscan
    type: exact
    k_values: [10]
    max_queries: 200
    sweep_param: none
    sweep_values: [0]

concurrency: [1, 4, 16]

churn:
  enabled: false

sift1m:
  enabled: false
