Summary

Metric	Model A	Model B	Winner

Labelled quality

Latency distribution (single + batch)

Single-query (warm) p50 / p95 / p99 (ms)

Batch latency p50 / p95 / p99 by concurrency (ms)

Throughput vs concurrency

Storage

Query-level detail

Show top-5 results per model side-by-side ({{QUERY_COUNT}} queries; click a query to expand)

Recommendation

The picker below uses fixed thresholds (no machine-learned scorer). Each axis has an independent winner; the global "no clear winner" verdict means axes disagree. Pick the model that wins your highest-weighted axis.

Disclaimers

This report is your-KB selection guidance, NOT an MTEB leaderboard (RFC 013 §3.2 N10).
Latency depends on hardware; rerun on the deployment target.
Labelled quality metrics require --golden; absent them, only fixture recall and Jaccard / Spearman diagnostics are shown.
Cost numbers reflect src/cost-estimates.ts LAST_VERIFIED={{COST_LAST_VERIFIED}}; verify if stale.
One machine, one moment. Comparing across hardware = run the orchestrator on each and diff manually.

{{TITLE}}