Benchmarks & Positioning

All numbers below are from frozen regression datasets, measured end-to-end against the running engine (re-baseline 2026-09-02, release build). Dev and heldout splits are kept separate; nothing is tuned against heldout.


Against a Production Baseline

The Argus-faithful dataset (200 production-shaped surveillance records: telegraphic enums, NA sparsity) compares PrismDB against the production system’s substring comparator under a protocol deliberately generous to the baseline (free keyword decomposition, precision-rewarding scoring, corpus-aware queries):

SystemScore
PrismDB37/50
Production substring baseline34/50

The natural-language suite is where the gap opens (15/25, with age-vocabulary queries won via automatic enum projection: 60-plus → elderly, 20s → young). Zero manual aliases were declared; the automatic tiers covered everything. The baseline’s failures are a structural ceiling (substring matching cannot bridge vocabulary); PrismDB’s remaining failures have identified mechanisms and roadmaps.


Regression Suites

DatasetShapeScore
Catalog30 records, 19 queries19/19
Large catalog250 records, 56 queries (dev/heldout)52/56 (26/28 + 26/28)
Surveillance proxy200 vision-LLM records, 12 dims30/50
Argus-faithful200 production-shaped records37/50 vs baseline 34/50
Argus v2 (mixed entity)250 records, 50 queries (25 dims)34/50 vs baseline 42/50
Workspace temporal15 tracks, 51 timestamped chunks11/16

Remaining workspace failures are documented honest boundaries: NLI correctly refusing weak-evidence matches, one measured dense-prefilter limit, one cross-track re-identification query that is out of scope.


Honest Positioning

Where PrismDB wins: attribute-decomposed queries over structured records, meaning queries whose terms map to independent record fields ("blue leather shoes", "elderly person carrying a red bag"). Per-dimension routing plus write-time vocabulary expansion is structurally better there than substring matching or one monolithic embedding.

Known structural boundaries:

  • Many-dimension corpora: on Argus v2 (25 dims, mixed person/vehicle records) the cross-dimension gate drops true matches that the substring baseline keeps (34/50 vs 42/50). This is the current engine round.
  • Reranker: a cross-encoder reranker over the same candidates is a rung on the cost ladder, weighed like every other tier; not yet measured on a fit judge.
  • Interpretive leaps invisible to dense retrieval (portfolio → document): beyond the current lexicon-projection tier’s reach; the query-log learning tier is the identified path.

Next public benchmark, STaRK-Amazon (NeurIPS 2024, planned): multi-attribute conversational product queries over structured entities (1.03M products), exactly the attribute-decomposed shape PrismDB is built for. Published dense baseline is 39.16 Hit@1 with LLM-based SOTA at 61.73 on the full corpus; PrismDB will be measured on a small subset with local baselines (substring, BM25, dense) on the same subset, so its numbers are not comparable to the published ones. A benchmark only counts here if it fits: the query shape has to be attribute-decomposed over structured records, the baselines have to run on the same subset, and the split has to be fixed before anything is measured. A suite that fails that test is not published as a result.

Every score above comes from a frozen dataset and a fixed protocol, re-measured end to end against the running engine. The datasets and the harness travel with the source, so the runs are reproducible for anyone holding it.