Benchmarks
measured retrieval, latency, memory, index size and cost evidence
Evidence is published with identity and context. A missing measurement is displayed as missing rather than replaced with an estimate.
Current service baseline
| kind | observational |
|---|---|
| measured | 2026-08-06 |
| code commit | 34461dc |
| live corpus commits | LU 65bc005, EU 6974532 |
| sampled MCP requests, 7 days | 0 |
| internal latency | p50 6.94 ms, p95 8.84 ms, p99 25.10 ms |
| average working set | 114 MiB |
Application Insights aggregates describe the current service, not the future hybrid retrieval path.
Retrieval relevance
Not measured yet. The public 200-case suite is a gated milestone.
Hybrid will not become default until exact identifiers, temporal isolation, nDCG@10,
regression, latency and memory pass their recorded thresholds.
| case category | public judgments |
|---|---|
| bilingual | 30 |
| conceptual | 60 |
| exact | 30 |
| fuzzy | 20 |
| hierarchy | 20 |
| temporal | 40 |
Download all 200 public cases and judgments. Each case names its language, time scope, relevant work, explanation and review status.
Publication rule
Future reports name code and corpus commits, the signed artifact manifest, embedding model, machine or Azure resource, timestamp, sample count and review status. Initial relevance judgments are labelled engineer-reviewed until a lawyer reviews them.