Jonex
EN 中文
Try Jonex FAQ

RESULTS & METHODOLOGY

Measured on reasoning. Verified by evidence.

Each evaluation answers a different question, so every number keeps its proper meaning.

MULTI-HOP REASONING

Five of six benchmark metrics improved.

Jonex + Jonex-Wiki is compared with LightRAG across 500 questions in each benchmark. The HotpotQA EM regression remains visible.

LightRAG and Jonex scores by benchmark and metric
BenchmarkMetricLightRAGJonex + Jonex-WikiDelta
HotpotQAF181.9%88.2%+6.3
HotpotQAEM68.2%66.7%-1.5
MuSiQueF165.9%82.1%+16.2
MuSiQueEM55.0%71.4%+16.4
2WikiMHQAF184.7%89.1%+4.4
2WikiMHQAEM76.4%77.1%+0.7

Scores use a full 0–100 scale. F1 and EM follow each benchmark protocol; baseline snapshot 2026-05.

FINANCEBENCH EVIDENCE-CHAIN REPLAY

Every replay completed and located its annotated evidence.

This public sample tests whether the knowledge workflow can recover and cite evidence marked in financial reports.

Questions
150
Evidence sets located
150/150
Evidence coverage
1.0
This is an evidence-chain replay, not end-to-end FinanceBench answer accuracy.

ANONYMIZED ENTERPRISE REPLAY

Answers remained accurate and traceable in internal historical scenarios.

De-identified policy and process questions were replayed across real enterprise materials.

Answer accuracy
96%
Source traceability
100%
Internal de-identified historical scenario replay. Scope and evaluation boundaries apply.