feat(admin): 「自动评测」页做实 —— 接真评测数据,去 mock (P1)

审计 P1「admin 三页纯 mock」之一。此前 EvalsPage 是写死的质量趋势+编造的
错题本+虚构纠偏轨迹。现接 sundynix_eval 真数据(评测经 JetStream eval 流持久
落库,刚升级)。

- store/eval_query.go:EvalTrend(按天 avg 综合分/忠实度+低分计数)、EvalSummaryFor
  (ok/warn/poor/corrected 计数+均值)、PoorEvals(错题本,level in poor/warn +
  评语+纠偏标记+租户名)。全 WithoutTenant 平台口径;忠实度均值只算 sources>0
  (无来源的忠实度恒0会压低失真)。
- GET /admin/evals?days=(RequireAdmin);admin api.ts + EvalsPage 重写:
  总览卡片(综合分/合格率/低分占比/纠偏采纳率)+质量&忠实度趋势(纯SVG折线+低分
  背景条)+错题本(点行展开评语)。
- 诚实边界:纠偏「前后全文对照」后端未持久化,只存了 Reason/Corrected/各维度分,
  故错题本展示评语+「已纠偏」标记,不再编造 before/after。

live:/admin/evals 返 57 次评测 avg=0.88、错题本10条、11天趋势;浏览器渲染
真数据(趋势线07-15真实下探)。go+tsc+41 vitest 全绿。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Blizzard
2026-07-18 12:03:54 +08:00
parent a1c35852ef
commit d04d830c37
5 changed files with 329 additions and 241 deletions
+44
View File
@@ -301,6 +301,50 @@ export async function adminReconcile(): Promise<{ diffs: ReconcileDiff[]; ok: bo
return { diffs: d.diffs ?? [], ok: !!d.ok };
}
// ---- 自动评测观测(真数据,来自 sundynix_eval----
export interface EvalDay {
day: string; // YYYYMMDD
avg_overall: number;
avg_faithful: number;
count: number;
poor_count: number;
}
export interface EvalSummary {
total: number;
ok: number;
warn: number;
poor: number;
corrected: number;
avg_overall: number;
}
export interface PoorEval {
task_id: string;
tenant_name: string;
owner: string;
overall: number;
rule: number;
llm: number;
faithful: number;
level: string;
reason: string;
sources: number;
corrected: boolean;
created_at: string;
}
export async function adminEvals(days = 14): Promise<{ from: string; to: string; trend: EvalDay[]; summary: EvalSummary; poor: PoorEval[] }> {
const res = guard(await fetch(`${ADMIN}/evals?days=${days}`, { headers: authHeaders() }));
const d = (await res.json().catch(() => ({}))) as { from?: string; to?: string; trend?: EvalDay[]; summary?: EvalSummary; poor?: PoorEval[]; error?: string };
if (!res.ok) throw new Error(d.error ?? `evals failed: ${res.status}`);
return {
from: d.from ?? "",
to: d.to ?? "",
trend: d.trend ?? [],
summary: d.summary ?? { total: 0, ok: 0, warn: 0, poor: 0, corrected: 0, avg_overall: 0 },
poor: d.poor ?? [],
};
}
// gatewayOnline 用公开的 /healthz 探活(不受鉴权影响)。
export async function gatewayOnline(): Promise<boolean> {
try {