Search Quality Benchmark
Overall quality
Mean relative shows average quality across the selected dataset/query-set cohorts. Quality floor shows the engine’s worst relative result: how close it stays to the best observed result even on its weakest cohort. Engines missing a selected cohort remain visible but are not ranked.
Quality consistency across datasets
Each small dot is one selected dataset, normalized to the best observed result on that dataset. The ring marks the engine’s weakest dataset (Quality floor); the diamond marks Mean relative. A tight cluster near 100% means consistently strong quality without large workload-specific drops. Engines with partial coverage are labeled with their dataset count and use a hollow mean marker.
Quality × latency
Use latency to understand the cost of search quality on this dataset. Performance is supporting context only and does not affect benchmark ranking.
Dataset × engine quality matrix
Each cell uses the best result among the selected search modes for one dataset/query-set cohort. TREC-COVID appears once per official topic field (query, question, narrative); distinct short and long query sets are separate columns, and identical texts share one “same text” cohort. “Per-dataset relative” is the engine score divided by the best observed score in that cohort. Color uses a fixed scale, so the same percentage always has the same color across filters and screenshots.
Search mode comparison
Compares Full-text, Vector, and Hybrid on one normalized cross-dataset scale. Each point is that engine/mode’s mean per-dataset quality relative to the best observed result in the same dataset. Lines connect the modes in the order Full-text → Vector → Hybrid.
Latency shape
The three points show typical latency (P50), slower requests (P95), and tail latency (P99). A short line means consistent response times. The horizontal scale is logarithmic; lower is better.
Slow-query inspection
Eight independent slowest successful query samples from every run matching the focused dataset, engine and search-mode filters, ordered by latency.
Slow-query inspection
Eight independent slowest successful query samples from every run matching the focused dataset, engine and search-mode filters, ordered by latency.
Run explorer
Inspect the underlying measured runs. Quality is the benchmark target; latency and QPS are included as supporting context. The filters above remain active.