Deterministic CC0 corpus

Million-Chunk Performance

Paired query latency, controlled indexing, full-system paths, and pinned public retrieval quality.

Warm p95

3.57x

53.77 ms -> 15.07 ms

Index throughput

21.96x

109006 chunks/s

Footprint

-57.0%

1.06 GiB -> 0.46 GiB

Acceptance

PASS

Latency, quality, footprint, recall, and documented indexing ceiling.

Full-system query paths

PathBaseline p95 msCurrent p95 msRatio
Process cold307.22478.021.556
Warm distinct112.5758.290.518
Cache replay27.8531.961.148
Filtered257.3145.220.176
Warm CLI151.27138.990.919
Concurrent164.7499.170.602

Public retrieval quality

MetricBaselineCurrentDelta
ndcg_at_100.2620 +/- 0.00020.2666 +/- 0.0016+0.0046
mrr_at_100.2178 +/- 0.00040.2220 +/- 0.0014+0.0042
precision_at_50.0561 +/- 0.00020.0601 +/- 0.0015+0.0039
recall_at_200.5080 +/- 0.00240.4890 +/- 0.0008-0.0190
no_hit_rate0.0000 +/- 0.00000.0000 +/- 0.0000+0.0000

Per-dataset nDCG@10

DatasetBaselineCurrentDelta
codetrans-dl0.2365 +/- 0.00050.2402 +/- 0.0033+0.0037
codetrans-contest0.4269 +/- 0.00000.4278 +/- 0.0001+0.0009
cosqa0.1464 +/- 0.00030.1566 +/- 0.0041+0.0102
codefeedback-st0.5243 +/- 0.00000.5108 +/- 0.0000-0.0135

Method

The query run kept both daemons live and alternated request order. The controlled indexing run used the same generated corpus for both binaries. The exact full-system run is retained separately because the shared host was heavily saturated.