feat: 支持 Agent 并发批测与洞察导出

2026-08-13 17:25:22 +08:00
parent 3faecab6ac
commit f952d6dc58
22 changed files with 666 additions and 44 deletions
--- a/ai_knowledge_base_v2/.env.example
+++ b/ai_knowledge_base_v2/.env.example
@@ -35,6 +35,8 @@ TOPIC_SETTLEMENT_MAX_ATTEMPTS=3
 AGENT_BATCH_WORKER_ENABLED=true
 AGENT_BATCH_POLL_SECONDS=2
 AGENT_BATCH_STALE_MINUTES=30
+# 单个后端进程同时执行的批量测试问题上限；任务自身还受页面选择的 2-10 路限制。
+AGENT_BATCH_WORKER_CONCURRENCY=10

 # ============================================
 # 内网穿透（frpc sidecar）
--- a/ai_knowledge_base_v2/apps/admin-web/src/components/AgentBatchTestPanel.vue
+++ b/ai_knowledge_base_v2/apps/admin-web/src/components/AgentBatchTestPanel.vue
@@ -15,6 +15,8 @@ const props = defineProps<{
 const fileInput = ref<HTMLInputElement | null>(null);
 const selectedFile = ref<File | null>(null);
 const taskName = ref("");
+const executionMode = ref<"sequential" | "concurrent">("sequential");
+const concurrencyLimit = ref(5);
 const creating = ref(false);
 const loading = ref(false);
 const exportingId = ref<number | null>(null);
@@ -81,7 +83,12 @@ async function createJob() {
  if (!String(props.config.modelId || "")) return ElMessage.warning("请先在调试配置中选择模型");
  creating.value = true;
  try {
-    const created = await api.createAgentBatch(selectedFile.value, { ...props.config, name: taskName.value.trim() });
+    const created = await api.createAgentBatch(selectedFile.value, {
+      ...props.config,
+      name: taskName.value.trim(),
+      executionMode: executionMode.value,
+      concurrencyLimit: executionMode.value === "sequential" ? 1 : concurrencyLimit.value,
+    });
    ElMessage.success(`已创建批量测试任务，共 ${created.totalCount} 个问题`);
    clearFile();
    taskName.value = "";
@@ -154,6 +161,10 @@ function statusType(status: string) {
  return "warning";
 }

+function executionLabel(job: AgentBatchJob) {
+  return job.executionMode === "concurrent" ? `并发 · ${job.concurrencyLimit} 路` : "逐条串行";
+}
+
 function formatDate(value?: string | null) {
  return value ? new Date(value).toLocaleString("zh-CN", { hour12: false }) : "—";
 }
@@ -170,7 +181,7 @@ function errorMessage(error: unknown, fallback: string) {
 <template>
  <div class="agent-batch-panel">
    <div class="agent-section-intro">
-      <div><h3>Excel 批量测试</h3><p>使用当前编辑框中的提示词、调试模型、知识库和生成参数创建配置快照，后台逐条生成答案。</p></div>
+      <div><h3>Excel 批量测试</h3><p>普通管理员仅看到自己的任务，超级管理员可统一审计；创建时会固化当前提示词、模型、知识库和生成参数。</p></div>
      <el-button @click="downloadTemplate">下载导入模板</el-button>
    </div>

@@ -180,6 +191,14 @@ function errorMessage(error: unknown, fallback: string) {
        <span><small>知识库</small><strong>{{ knowledgeSummary }}</strong></span>
        <span><small>单次上限</small><strong>200 个问题</strong></span>
      </div>
+      <div class="agent-batch-mode-setting">
+        <div class="agent-batch-mode-copy"><strong>执行方式</strong><span>串行适合日常回归；并发适合压测吞吐与稳定性，并发数会被系统安全上限约束。</span></div>
+        <el-radio-group v-model="executionMode" class="agent-batch-mode-options">
+          <el-radio-button value="sequential">逐条串行</el-radio-button>
+          <el-radio-button value="concurrent">并发测试</el-radio-button>
+        </el-radio-group>
+        <label v-if="executionMode === 'concurrent'" class="agent-batch-concurrency"><span>并发数</span><el-input-number v-model="concurrencyLimit" :min="2" :max="10" controls-position="right" /></label>
+      </div>
      <div class="agent-batch-upload-row">
        <label class="agent-batch-name"><span>任务名称</span><input v-model="taskName" maxlength="160" placeholder="例如：8月主提示词回归测试" /></label>
        <label class="agent-batch-file" :class="{ selected: selectedFile }">
@@ -197,6 +216,7 @@ function errorMessage(error: unknown, fallback: string) {
      <el-table-column label="任务" min-width="190"><template #default="{ row }"><strong>{{ row.name }}</strong><small class="agent-batch-cell-note">#{{ row.id }} · {{ formatDate(row.createdAt) }}</small></template></el-table-column>
      <el-table-column label="配置快照" min-width="180"><template #default="{ row }"><span>{{ row.modelName }}</span><small class="agent-batch-cell-note">{{ row.knowledgeNames.join('、') }}</small></template></el-table-column>
      <el-table-column label="进度" min-width="190"><template #default="{ row }"><el-progress :percentage="row.progress" :status="row.status === 'completed' ? 'success' : undefined" /><small class="agent-batch-cell-note">{{ row.processedCount }}/{{ row.totalCount }} · 成功 {{ row.successCount }} · 失败 {{ row.failedCount }}</small></template></el-table-column>
+      <el-table-column label="执行方式" width="118"><template #default="{ row }"><el-tag type="info">{{ executionLabel(row) }}</el-tag></template></el-table-column>
      <el-table-column label="状态" width="108"><template #default="{ row }"><el-tag :type="statusType(row.status)">{{ statusLabel(row.status) }}</el-tag></template></el-table-column>
      <el-table-column label="操作" width="190" fixed="right"><template #default="{ row }"><el-button link type="primary" @click="openDetail(row)">查看</el-button><el-button link type="primary" :loading="exportingId === row.id" @click="exportJob(row)">导出</el-button><el-button v-if="['pending','running'].includes(row.status)" link type="danger" @click="cancelJob(row)">取消</el-button></template></el-table-column>
    </el-table>
@@ -207,7 +227,7 @@ function errorMessage(error: unknown, fallback: string) {
    <div v-loading="detailLoading" class="agent-batch-detail">
      <template v-if="selectedJob">
        <div class="agent-batch-detail-head">
-          <div><strong>{{ selectedJob.name }}</strong><span>{{ selectedJob.modelName }} · {{ selectedJob.knowledgeNames.join('、') }}</span></div>
+          <div><strong>{{ selectedJob.name }}</strong><span>{{ selectedJob.modelName }} · {{ selectedJob.knowledgeNames.join('、') }} · {{ executionLabel(selectedJob) }}</span></div>
          <el-tag :type="statusType(selectedJob.status)">{{ statusLabel(selectedJob.status) }}</el-tag>
        </div>
        <el-table :data="selectedJob.items.items" stripe max-height="520">
--- a/ai_knowledge_base_v2/apps/admin-web/src/components/AgentManagementView.vue
+++ b/ai_knowledge_base_v2/apps/admin-web/src/components/AgentManagementView.vue
@@ -122,7 +122,7 @@ async function load() {
  loading.value = true;
  try {
    const [promptResult, modelRows, knowledgeRows, formalConfig] = await Promise.all([
-      api.prompt(), api.models(), api.knowledgeOptions(), api.agentRuntimeConfig(),
+      api.prompt(), api.agentModelOptions(), api.agentKnowledgeOptions(), api.agentRuntimeConfig(),
    ]);
    applyPrompt(promptResult);
    models.value = modelRows;
--- a/ai_knowledge_base_v2/apps/admin-web/src/components/RecordAuditView.vue
+++ b/ai_knowledge_base_v2/apps/admin-web/src/components/RecordAuditView.vue
@@ -14,6 +14,7 @@ import ChatDetailDrawer from "./ChatDetailDrawer.vue";

 type RecordTab = "chats" | "questionInsights" | "aiLogs" | "operationLogs";
 const loading = ref(false);
+const insightExporting = ref(false);
 const activeTab = ref<RecordTab>("chats");
 const chats = ref<ChatRecord[]>([]);
 const ssoClients = ref<SsoClientItem[]>([]);
@@ -175,6 +176,32 @@ async function resetInsightFilters() {
  });
  await refreshInsights();
 }
+async function exportInsights() {
+  const dateFrom = formatDateTime(insightFilters.dateFrom, "start");
+  const dateTo = formatDateTime(insightFilters.dateTo, "end");
+  if (dateFrom && dateTo && new Date(dateFrom.replace(" ", "T")) > new Date(dateTo.replace(" ", "T"))) {
+    return ElMessage.warning("开始时间不能晚于结束时间");
+  }
+  insightExporting.value = true;
+  try {
+    const fromLabel = insightFilters.dateFrom.slice(0, 10).replaceAll("-", "") || "全部";
+    const toLabel = insightFilters.dateTo.slice(0, 10).replaceAll("-", "") || "全部";
+    await api.exportQuestionInsights(
+      {
+        dateFrom,
+        dateTo,
+        minCount: insightFilters.minCount,
+        maxMessages: insightFilters.maxMessages,
+      },
+      `问题洞察_${fromLabel}_${toLabel}.xlsx`,
+    );
+    ElMessage.success("问题洞察已导出");
+  } catch (error) {
+    ElMessage.error(error instanceof Error ? error.message : "问题洞察导出失败");
+  } finally {
+    insightExporting.value = false;
+  }
+}
 function openChat(sessionId: number) {
  detailSessionId.value = sessionId;
  chatDetailOpen.value = true;
@@ -355,6 +382,7 @@ function formatMoney(value?: number | null, currency = "CNY") {
              @click="refreshInsights"
              >刷新并统计</el-button
            ><el-button @click="resetInsightFilters">重置</el-button>
+            <el-button :loading="insightExporting" @click="exportInsights">导出洞察</el-button>
          </div>
        </div>
        <p class="question-insight-help">
--- a/ai_knowledge_base_v2/apps/admin-web/src/main.ts
+++ b/ai_knowledge_base_v2/apps/admin-web/src/main.ts
@@ -20,6 +20,8 @@ import {
  ElOption,
  ElPagination,
  ElProgress,
+  ElRadioButton,
+  ElRadioGroup,
  ElSelect,
  ElSlider,
  ElSwitch,
@@ -54,6 +56,8 @@ import "element-plus/theme-chalk/el-option.css";
 import "element-plus/theme-chalk/el-overlay.css";
 import "element-plus/theme-chalk/el-pagination.css";
 import "element-plus/theme-chalk/el-progress.css";
+import "element-plus/theme-chalk/el-radio-button.css";
+import "element-plus/theme-chalk/el-radio-group.css";
 import "element-plus/theme-chalk/el-popper.css";
 import "element-plus/theme-chalk/el-scrollbar.css";
 import "element-plus/theme-chalk/el-select.css";
@@ -95,6 +99,8 @@ app.config.globalProperties.$ELEMENT = { locale: zhCn };
  ElOption,
  ElPagination,
  ElProgress,
+  ElRadioButton,
+  ElRadioGroup,
  ElSelect,
  ElSlider,
  ElSwitch,
@@ -108,6 +114,9 @@ app.config.globalProperties.$ELEMENT = { locale: zhCn };
 ].forEach((component) => {
  app.use(component);
 });
+// Element Plus 2.11 exposes radio group/button without an installer; register them explicitly.
+app.component("ElRadioButton", ElRadioButton);
+app.component("ElRadioGroup", ElRadioGroup);

 app.directive("loading", vLoading);
 app.mount("#app");
--- a/ai_knowledge_base_v2/apps/admin-web/src/services/api.ts
+++ b/ai_knowledge_base_v2/apps/admin-web/src/services/api.ts
@@ -261,6 +261,8 @@ export const api = {
  agentRuntimeConfig: () => request<AgentRuntimeConfig>("/admin/agent/runtime-config"),
  saveAgentRuntimeConfig: (payload: AgentGenerationConfig) =>
    request<AgentRuntimeConfig>("/admin/agent/runtime-config", { method: "PUT", body: JSON.stringify(payload) }),
+  agentKnowledgeOptions: () => request<KnowledgeItem[]>("/admin/agent/knowledge-options"),
+  agentModelOptions: () => request<ModelItem[]>("/admin/agent/model-options"),
  downloadAgentBatchTemplate: () => download("/admin/agent/batch/template", "Agent批量测试导入模板.xlsx"),
  createAgentBatch: (file: File, config: Record<string, unknown>) => {
    const formData = new FormData();
@@ -334,6 +336,8 @@ export const api = {
    request<QuestionInsightRefreshResult>(`/admin/question-insights/refresh${queryString(query)}`, {
      method: "POST",
    }),
+  exportQuestionInsights: (query: { dateFrom?: string; dateTo?: string; minCount?: number; maxMessages?: number } = {}, filename = "问题洞察.xlsx") =>
+    download(`/admin/question-insights/export${queryString(query)}`, filename),
  retrievalLogs: (query: { page?: number; pageSize?: number } = {}) => request<PageResult<RetrievalLogItem>>(`/admin/retrieval-log/list${queryString(query)}`),
  estimateRetrievalCleanup: (before: string) => request<{ before: string; estimatedCount: number }>("/admin/retrieval-log/cleanup/estimate", { method: "POST", body: JSON.stringify({ before }) }),
  cleanupRetrievalLogs: (before: string) => request<{ before: string; deleted: number }>("/admin/retrieval-log/cleanup", { method: "POST", body: JSON.stringify({ before }) }),
--- a/ai_knowledge_base_v2/apps/admin-web/src/styles.css
+++ b/ai_knowledge_base_v2/apps/admin-web/src/styles.css
@@ -1061,6 +1061,14 @@ textarea {
 .agent-batch-config-summary small { margin-bottom: 4px; color: #7a8c85; font-size: 11px; }
 .agent-batch-config-summary strong { overflow: hidden; color: #2b4038; font-size: 13px; text-overflow: ellipsis; white-space: nowrap; }
 .agent-batch-upload-row { display: flex; align-items: flex-end; gap: 10px; flex-wrap: wrap; }
+.agent-batch-mode-setting { display: flex; align-items: center; gap: 14px; padding: 12px; border: 1px solid #dfe8e5; border-radius: 9px; background: #fff; }
+.agent-batch-mode-copy { display: grid; min-width: 240px; flex: 1; gap: 3px; }
+.agent-batch-mode-copy strong { color: #2b4038; font-size: 13px; }
+.agent-batch-mode-copy span { color: #71827c; font-size: 11px; line-height: 1.5; }
+.agent-batch-mode-options { flex: none; }
+.agent-batch-concurrency { display: flex; align-items: center; gap: 8px; color: #536860; font-size: 12px; }
+.agent-batch-concurrency > span { white-space: nowrap; }
+.agent-batch-concurrency .el-input-number { width: 112px; }
 .agent-batch-name { display: grid; gap: 6px; width: 220px; color: #536860; font-size: 12px; }
 .agent-batch-name input { height: 40px; padding: 0 12px; border: 1px solid #d8e2df; border-radius: 6px; outline: none; background: #fff; color: #263832; }
 .agent-batch-name input:focus { border-color: #2f9479; box-shadow: 0 0 0 2px rgba(47, 148, 121, .12); }
@@ -2349,6 +2357,11 @@ textarea {

  .agent-batch-config-summary { grid-template-columns: 1fr; }
  .agent-batch-name, .agent-batch-file { width: 100%; min-width: 0; }
+  .agent-batch-mode-setting { align-items: stretch; flex-direction: column; }
+  .agent-batch-mode-copy { min-width: 0; }
+  .agent-batch-mode-options { display: flex; }
+  .agent-batch-mode-options .el-radio-button { flex: 1; }
+  .agent-batch-mode-options .el-radio-button__inner { width: 100%; }
  .agent-batch-upload-row > .el-button { flex: 1; }
 }

--- a/ai_knowledge_base_v2/apps/admin-web/src/types/api.ts
+++ b/ai_knowledge_base_v2/apps/admin-web/src/types/api.ts
@@ -404,6 +404,8 @@ export interface AgentBatchJob {
  successCount: number;
  failedCount: number;
  progress: number;
+  executionMode: "sequential" | "concurrent";
+  concurrencyLimit: number;
  modelId: number;
  modelName: string;
  knowledgeIds: number[];
--- a/ai_knowledge_base_v2/apps/backend/.env.example
+++ b/ai_knowledge_base_v2/apps/backend/.env.example
@@ -53,6 +53,10 @@ TOPIC_SETTLEMENT_WORKER_ENABLED=true
 TOPIC_SETTLEMENT_POLL_SECONDS=2
 TOPIC_SETTLEMENT_STALE_MINUTES=30
 TOPIC_SETTLEMENT_MAX_ATTEMPTS=3
+AGENT_BATCH_WORKER_ENABLED=true
+AGENT_BATCH_POLL_SECONDS=2
+AGENT_BATCH_STALE_MINUTES=30
+AGENT_BATCH_WORKER_CONCURRENCY=10

 # 本地开发可以使用开发密码；生产环境必须改成高强度密码，且 APP_ENV=production 时不能使用 admin123456
 BOOTSTRAP_ADMIN_USERNAME=admin
--- a/ai_knowledge_base_v2/apps/backend/alembic/versions/0034_agent_batch_concurrency.py
+++ b/ai_knowledge_base_v2/apps/backend/alembic/versions/0034_agent_batch_concurrency.py
@@ -0,0 +1,30 @@
+"""add Agent batch execution mode and concurrency limit
+
+Revision ID: 0034_agent_batch_concurrency
+Revises: 0033_agent_batch_tests
+"""
+
+from alembic import op
+import sqlalchemy as sa
+
+
+revision = "0034_agent_batch_concurrency"
+down_revision = "0033_agent_batch_tests"
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "sys_agent_batch_test",
+        sa.Column("execution_mode", sa.String(20), nullable=False, server_default="sequential"),
+    )
+    op.add_column(
+        "sys_agent_batch_test",
+        sa.Column("concurrency_limit", sa.Integer(), nullable=False, server_default="1"),
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("sys_agent_batch_test", "concurrency_limit")
+    op.drop_column("sys_agent_batch_test", "execution_mode")
--- a/ai_knowledge_base_v2/apps/backend/app/api/admin_agent_batch.py
+++ b/ai_knowledge_base_v2/apps/backend/app/api/admin_agent_batch.py
@@ -16,6 +16,8 @@ from app.core.dependencies import get_current_admin
 from app.core.responses import api_success
 from app.models.admin import Admin
 from app.models.agent_batch import AgentBatchTest, AgentBatchTestItem
+from app.models.ai_config import ModelConfig
+from app.models.knowledge import Knowledge
 from app.services.admin_service import OperationLogService
 from app.services.agent_batch_test_service import AgentBatchTestService, item_dict, job_dict

@@ -23,6 +25,58 @@ from app.services.agent_batch_test_service import AgentBatchTestService, item_di
 router = APIRouter()


+@router.get("/agent/model-options")
+def agent_model_options(
+    db: Session = Depends(get_db),
+    current_admin: Admin = Depends(get_current_admin),
+) -> dict:
+    """Safe Agent-only model choices without provider credentials or endpoints."""
+    items = list(db.scalars(select(ModelConfig).order_by(ModelConfig.is_default.desc(), ModelConfig.id.desc())))
+    return api_success([
+        {
+            "id": item.id,
+            "displayName": item.display_name,
+            "modelName": item.model_name,
+            "temperature": float(item.temperature) if item.temperature is not None else None,
+            "topP": float(item.top_p) if item.top_p is not None else None,
+            "topK": item.top_k,
+            "presencePenalty": float(item.presence_penalty) if item.presence_penalty is not None else None,
+            "frequencyPenalty": float(item.frequency_penalty) if item.frequency_penalty is not None else None,
+            "maxToken": item.max_token,
+            "streamEnabled": item.stream_enabled,
+            "enabled": item.enabled,
+            "isDefault": item.is_default,
+        }
+        for item in items
+    ])
+
+
+@router.get("/agent/knowledge-options")
+def agent_knowledge_options(
+    db: Session = Depends(get_db),
+    current_admin: Admin = Depends(get_current_admin),
+) -> dict:
+    """Safe Agent-only knowledge choices; does not grant Knowledge Management access."""
+    items = list(
+        db.scalars(
+            select(Knowledge)
+            .where(Knowledge.lifecycle_status == "active")
+            .order_by(Knowledge.id.desc())
+        )
+    )
+    return api_success([
+        {
+            "id": item.id,
+            "name": item.name,
+            "status": item.status,
+            "knowledgeType": item.knowledge_type,
+            "lifecycleStatus": item.lifecycle_status,
+            "sourceStatus": item.source_status,
+        }
+        for item in items
+    ])
+
+
@router.get("/agent/batch/template")
 def download_batch_template(
    current_admin: Admin = Depends(get_current_admin),
@@ -67,10 +121,12 @@ def list_batch_jobs(
    db: Session = Depends(get_db),
    current_admin: Admin = Depends(get_current_admin),
 ) -> dict:
-    total = db.scalar(select(func.count(AgentBatchTest.id))) or 0
+    owner_filter = _owner_filter(current_admin)
+    total = db.scalar(select(func.count(AgentBatchTest.id)).where(*owner_filter)) or 0
    rows = list(
        db.scalars(
            select(AgentBatchTest)
+            .where(*owner_filter)
            .order_by(AgentBatchTest.created_at.desc(), AgentBatchTest.id.desc())
            .offset((page - 1) * pageSize)
            .limit(pageSize)
@@ -87,7 +143,7 @@ def get_batch_job(
    db: Session = Depends(get_db),
    current_admin: Admin = Depends(get_current_admin),
 ) -> dict:
-    job = _job_or_404(db, job_id)
+    job = _job_or_404(db, job_id, current_admin)
    total = db.scalar(select(func.count(AgentBatchTestItem.id)).where(AgentBatchTestItem.batch_id == job.id)) or 0
    items = list(
        db.scalars(
@@ -107,7 +163,7 @@ def export_batch_job(
    db: Session = Depends(get_db),
    current_admin: Admin = Depends(get_current_admin),
 ) -> StreamingResponse:
-    job = _job_or_404(db, job_id)
+    job = _job_or_404(db, job_id, current_admin)
    items = list(
        db.scalars(
            select(AgentBatchTestItem)
@@ -131,7 +187,7 @@ def cancel_batch_job(
    db: Session = Depends(get_db),
    current_admin: Admin = Depends(get_current_admin),
 ) -> dict:
-    job = _job_or_404(db, job_id)
+    job = _job_or_404(db, job_id, current_admin)
    if job.status not in {"pending", "running"}:
        raise HTTPException(status_code=status.HTTP_409_CONFLICT, detail="当前任务已结束，不能取消")
    now = datetime.now(UTC).replace(tzinfo=None)
@@ -150,8 +206,12 @@ def cancel_batch_job(
    return api_success(job_dict(job))


-def _job_or_404(db: Session, job_id: int) -> AgentBatchTest:
+def _owner_filter(admin: Admin) -> tuple:
+    return () if admin.is_super_admin else (AgentBatchTest.created_by == admin.id,)
+
+
+def _job_or_404(db: Session, job_id: int, admin: Admin) -> AgentBatchTest:
    job = db.get(AgentBatchTest, job_id)
-    if job is None:
+    if job is None or (not admin.is_super_admin and job.created_by != admin.id):
        raise HTTPException(status_code=status.HTTP_404_NOT_FOUND, detail="批量测试任务不存在")
    return job
--- a/ai_knowledge_base_v2/apps/backend/app/api/admin_records.py
+++ b/ai_knowledge_base_v2/apps/backend/app/api/admin_records.py
@@ -7,7 +7,7 @@ from io import StringIO
 from typing import Any

 from fastapi import APIRouter, Depends, HTTPException, Query, status
-from fastapi.responses import Response
+from fastapi.responses import Response, StreamingResponse
 from sqlalchemy import exists, func, or_, select
 from sqlalchemy.orm import Session, load_only

@@ -23,6 +23,7 @@ from app.models.user import User
 from app.api.pagination import page_result
 from app.services.admin_service import OperationLogService
 from app.services.question_insight_service import QuestionInsightService
+from app.services.question_insight_export_service import QuestionInsightExportService
 from app.services.growth_profile_service import topic_dict, topic_summary_dict
 from app.services.help_card_service import help_card_dict
 from app.services.share_draft_service import share_draft_dict
@@ -324,6 +325,45 @@ def refresh_question_insights(
    return api_success(result)


+@router.get("/question-insights/export")
+def export_question_insights(
+    dateFrom: datetime | None = Query(default=None),
+    dateTo: datetime | None = Query(default=None),
+    minCount: int = Query(default=2, ge=1, le=50),
+    maxMessages: int = Query(default=5000, ge=100, le=20000),
+    db: Session = Depends(get_db),
+    current_admin: Admin = Depends(get_current_admin),
+) -> StreamingResponse:
+    if dateFrom is not None and dateTo is not None and dateFrom > dateTo:
+        raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="开始时间不能晚于结束时间")
+    result = QuestionInsightService.summarize(
+        db,
+        date_from=dateFrom,
+        date_to=dateTo,
+        min_count=minCount,
+        max_messages=maxMessages,
+        page=1,
+        page_size=maxMessages,
+    )
+    OperationLogService.write(
+        db,
+        admin_id=current_admin.id,
+        module="question_insight",
+        action=f"export:{result['total']}",
+    )
+    db.commit()
+    filename = f"question_insights_{_export_date_label(dateFrom)}_{_export_date_label(dateTo)}.xlsx"
+    return StreamingResponse(
+        QuestionInsightExportService.build(result),
+        media_type="application/vnd.openxmlformats-officedocument.spreadsheetml.sheet",
+        headers={"Content-Disposition": f'attachment; filename="{filename}"'},
+    )
+
+
+def _export_date_label(value: datetime | None) -> str:
+    return value.strftime("%Y%m%d") if value is not None else "all"
+
+
 def _chat_query(
    *,
    keyword: str,
--- a/ai_knowledge_base_v2/apps/backend/app/core/config.py
+++ b/ai_knowledge_base_v2/apps/backend/app/core/config.py
@@ -83,6 +83,7 @@ class Settings(BaseSettings):
    agent_batch_worker_enabled: bool = True
    agent_batch_poll_seconds: int = 2
    agent_batch_stale_minutes: int = 30
+    agent_batch_worker_concurrency: int = 10
    bootstrap_admin_username: str = ""
    bootstrap_admin_password: str = ""
    bootstrap_admin_name: str = "系统管理员"
--- a/ai_knowledge_base_v2/apps/backend/app/models/agent_batch.py
+++ b/ai_knowledge_base_v2/apps/backend/app/models/agent_batch.py
@@ -26,6 +26,8 @@ class AgentBatchTest(Base):
    success_count: Mapped[int] = mapped_column(Integer, default=0, nullable=False)
    failed_count: Mapped[int] = mapped_column(Integer, default=0, nullable=False)
    processed_count: Mapped[int] = mapped_column(Integer, default=0, nullable=False)
+    execution_mode: Mapped[str] = mapped_column(String(20), default="sequential", nullable=False)
+    concurrency_limit: Mapped[int] = mapped_column(Integer, default=1, nullable=False)
    prompt_content: Mapped[str] = mapped_column(Text, nullable=False)
    model_id: Mapped[int] = mapped_column(ForeignKey("sys_model.id"), nullable=False)
    model_name: Mapped[str] = mapped_column(String(100), nullable=False)
--- a/ai_knowledge_base_v2/apps/backend/app/services/agent_batch_test_service.py
+++ b/ai_knowledge_base_v2/apps/backend/app/services/agent_batch_test_service.py
@@ -19,6 +19,7 @@ from app.models.knowledge import Knowledge

 MAX_BATCH_QUESTIONS = 200
 MAX_QUESTION_LENGTH = 2000
+MAX_JOB_CONCURRENCY = 10
 TEMPLATE_HEADERS = ["序号", "问题"]


@@ -105,10 +106,25 @@ class AgentBatchTestService:
            "reasoningVisible": 0,
            "responseDepth": int(config.get("responseDepth", 35)),
        }
+        execution_mode = str(config.get("executionMode") or "sequential")
+        if execution_mode not in {"sequential", "concurrent"}:
+            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="批量测试执行模式不正确")
+        try:
+            requested_concurrency = int(config.get("concurrencyLimit") or 5)
+        except (TypeError, ValueError) as exc:
+            raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="并发数格式不正确") from exc
+        concurrency_limit = 1 if execution_mode == "sequential" else requested_concurrency
+        if execution_mode == "concurrent" and not 2 <= concurrency_limit <= MAX_JOB_CONCURRENCY:
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail=f"单个任务并发数必须在 2 到 {MAX_JOB_CONCURRENCY} 之间",
+            )
        job = AgentBatchTest(
            name=str(config.get("name") or filename.rsplit(".", 1)[0] or "Agent 批量测试")[:160],
            original_filename=filename[:255],
            total_count=len(rows),
+            execution_mode=execution_mode,
+            concurrency_limit=concurrency_limit,
            prompt_content=prompt_content,
            model_id=model.id,
            model_name=model.display_name or model.model_name,
@@ -183,13 +199,14 @@ class AgentBatchTestService:
        summary.append(["任务名称", job.name])
        summary.append(["任务状态", job.status])
        summary.append(["问题总数", job.total_count])
+        summary.append(["执行方式", "逐条串行" if job.execution_mode == "sequential" else f"并发测试（{job.concurrency_limit} 路）"])
        summary.append(["成功", job.success_count])
        summary.append(["失败", job.failed_count])
        summary.append(["测试模型", job.model_name])
        summary.append(["知识库", "、".join(json.loads(job.knowledge_names or "[]"))])
        summary.append(["创建时间", job.created_at])
        summary.append(["完成时间", job.finished_at])
-        _style_sheet(summary, widths=(20, 86), table_ref="A1:B10", table_name="AgentBatchSummary")
+        _style_sheet(summary, widths=(20, 86), table_ref="A1:B11", table_name="AgentBatchSummary")
        summary.column_dimensions["B"].width = 86
        stream = BytesIO()
        workbook.save(stream)
@@ -205,6 +222,8 @@ def job_dict(job: AgentBatchTest) -> dict:
        "status": job.status,
        "totalCount": job.total_count,
        "processedCount": job.processed_count,
+        "executionMode": job.execution_mode,
+        "concurrencyLimit": job.concurrency_limit,
        "successCount": job.success_count,
        "failedCount": job.failed_count,
        "progress": round(job.processed_count * 100 / job.total_count) if job.total_count else 0,
--- a/ai_knowledge_base_v2/apps/backend/app/services/agent_batch_test_worker.py
+++ b/ai_knowledge_base_v2/apps/backend/app/services/agent_batch_test_worker.py
@@ -33,13 +33,43 @@ class AgentBatchTestWorker:
            logger.info("agent batch test worker disabled")
            return
        worker_id = f"{socket.gethostname()}:{os.getpid()}:{uuid.uuid4().hex[:8]}"
+        active: set[asyncio.Task] = set()
+        worker_concurrency = max(1, min(settings.agent_batch_worker_concurrency, 50))
+        next_recovery_at = 0.0
        while True:
            try:
-                processed = await cls.run_once(worker_id)
+                if time.monotonic() >= next_recovery_at:
+                    with SessionLocal() as db:
+                        recovered_batch_ids = cls.recover_stale_items(db)
+                        db.commit()
+                        for batch_id in recovered_batch_ids:
+                            cls.refresh_job(db, batch_id)
+                    next_recovery_at = time.monotonic() + 60
+                active = {task for task in active if not task.done()}
+                claimed = False
+                while len(active) < worker_concurrency:
+                    with SessionLocal() as db:
+                        item_id = cls.claim_next(db, worker_id=worker_id)
+                    if item_id is None:
+                        break
+                    claimed = True
+                    task = asyncio.create_task(cls.execute_claimed(item_id=item_id, worker_id=worker_id))
+                    task.add_done_callback(cls._log_task_failure)
+                    active.add(task)
+                if active:
+                    await asyncio.wait(
+                        active,
+                        timeout=max(1, settings.agent_batch_poll_seconds),
+                        return_when=asyncio.FIRST_COMPLETED,
+                    )
+                    processed = True
+                else:
+                    processed = claimed
            except Exception:
                processed = False
                logger.exception("agent batch test worker iteration failed")
-            await asyncio.sleep(0 if processed else max(1, settings.agent_batch_poll_seconds))
+            if not processed:
+                await asyncio.sleep(max(1, settings.agent_batch_poll_seconds))

    @classmethod
    async def run_once(cls, worker_id: str) -> bool:
@@ -59,35 +89,64 @@ class AgentBatchTestWorker:
    @staticmethod
    def claim_next(db: Session, *, worker_id: str, now: datetime | None = None) -> int | None:
        current = now or _now()
-        item = db.scalar(
-            select(AgentBatchTestItem)
-            .join(AgentBatchTest, AgentBatchTest.id == AgentBatchTestItem.batch_id)
-            .where(
-                AgentBatchTest.status.in_(("pending", "running")),
-                AgentBatchTestItem.status == "pending",
-                AgentBatchTestItem.attempt_count < AgentBatchTestItem.max_attempts,
-                or_(AgentBatchTestItem.next_run_at.is_(None), AgentBatchTestItem.next_run_at <= current),
+        jobs = list(
+            db.scalars(
+                select(AgentBatchTest)
+                .where(AgentBatchTest.status.in_(("pending", "running")))
+                .order_by(AgentBatchTest.created_at.asc(), AgentBatchTest.id.asc())
+                .limit(50)
+                .with_for_update(skip_locked=True)
            )
-            .order_by(AgentBatchTestItem.batch_id.asc(), AgentBatchTestItem.row_number.asc())
-            .with_for_update(skip_locked=True)
-            .limit(1)
        )
-        if item is None:
-            db.rollback()
-            return None
-        job = db.get(AgentBatchTest, item.batch_id)
-        item.status = "running"
-        item.attempt_count += 1
-        item.locked_at = current
-        item.locked_by = worker_id
-        item.started_at = current
-        if job is not None and job.status == "pending":
-            job.status = "running"
-            job.started_at = current
-            db.add(job)
-        db.add(item)
-        db.commit()
-        return item.id
+        for job in jobs:
+            running_count = db.scalar(
+                select(func.count(AgentBatchTestItem.id)).where(
+                    AgentBatchTestItem.batch_id == job.id,
+                    AgentBatchTestItem.status == "running",
+                )
+            ) or 0
+            job_limit = 1 if job.execution_mode == "sequential" else max(1, job.concurrency_limit)
+            if running_count >= job_limit:
+                continue
+            item = db.scalar(
+                select(AgentBatchTestItem)
+                .where(
+                    AgentBatchTestItem.batch_id == job.id,
+                    AgentBatchTestItem.status == "pending",
+                    AgentBatchTestItem.attempt_count < AgentBatchTestItem.max_attempts,
+                    or_(AgentBatchTestItem.next_run_at.is_(None), AgentBatchTestItem.next_run_at <= current),
+                )
+                .order_by(AgentBatchTestItem.row_number.asc(), AgentBatchTestItem.id.asc())
+                .with_for_update(skip_locked=True)
+                .limit(1)
+            )
+            if item is None:
+                continue
+            item.status = "running"
+            item.attempt_count += 1
+            item.locked_at = current
+            item.locked_by = worker_id
+            item.started_at = current
+            if job.status == "pending":
+                job.status = "running"
+                job.started_at = current
+                db.add(job)
+            db.add(item)
+            db.commit()
+            return item.id
+        db.rollback()
+        return None
+
+    @staticmethod
+    def _log_task_failure(task: asyncio.Task) -> None:
+        if task.cancelled():
+            return
+        error = task.exception()
+        if error is not None:
+            logger.error(
+                "agent batch test execution task failed",
+                exc_info=(type(error), error, error.__traceback__),
+            )

    @classmethod
    async def execute_claimed(cls, *, item_id: int, worker_id: str) -> None:
--- a/ai_knowledge_base_v2/apps/backend/app/services/question_insight_export_service.py
+++ b/ai_knowledge_base_v2/apps/backend/app/services/question_insight_export_service.py
@@ -0,0 +1,167 @@
+from __future__ import annotations
+
+from datetime import datetime
+from io import BytesIO
+
+from openpyxl import Workbook
+from openpyxl.styles import Alignment, Font, PatternFill
+from openpyxl.worksheet.table import Table, TableStyleInfo
+
+
+class QuestionInsightExportService:
+    """Render a complete question-insight snapshot as an operator-friendly workbook."""
+
+    @staticmethod
+    def build(result: dict) -> BytesIO:
+        workbook = Workbook()
+        insight_sheet = workbook.active
+        insight_sheet.title = "洞察结果"
+        _append_insights(insight_sheet, result.get("items") or [])
+
+        sample_sheet = workbook.create_sheet("相似问法与样例")
+        _append_samples(sample_sheet, result.get("items") or [])
+
+        summary_sheet = workbook.create_sheet("统计说明")
+        _append_summary(summary_sheet, result)
+
+        stream = BytesIO()
+        workbook.save(stream)
+        stream.seek(0)
+        return stream
+
+
+def _append_insights(sheet, items: list[dict]) -> None:
+    headers = [
+        "排名", "问题组", "分类", "出现次数", "用户数", "会话数", "关联AI请求",
+        "无知识命中", "失败请求", "需补知识/查召回", "建议动作", "高频词",
+        "首次出现", "最后出现",
+    ]
+    sheet.append(headers)
+    for item in items:
+        sheet.append([
+            item.get("rank"),
+            _excel_safe(item.get("title")),
+            _excel_safe(item.get("categoryLabel")),
+            item.get("count", 0),
+            item.get("userCount", 0),
+            item.get("sessionCount", 0),
+            item.get("aiRequestCount", 0),
+            item.get("noHitCount", 0),
+            item.get("failedCount", 0),
+            "是" if item.get("needsKnowledgeFollowUp") else "否",
+            _excel_safe(item.get("suggestedAction")),
+            _excel_safe("、".join(item.get("topTerms") or [])),
+            _excel_datetime(item.get("firstSeenAt")),
+            _excel_datetime(item.get("lastSeenAt")),
+        ])
+    _style_table(
+        sheet,
+        widths=(8, 44, 14, 12, 10, 10, 14, 12, 12, 18, 54, 36, 20, 20),
+        table_name="QuestionInsightResults",
+    )
+    for row in sheet.iter_rows(min_row=2):
+        for index in (1, 10, 11):
+            row[index].alignment = Alignment(vertical="top", wrap_text=True)
+        for index in (12, 13):
+            row[index].number_format = "yyyy-mm-dd hh:mm:ss"
+
+
+def _append_samples(sheet, items: list[dict]) -> None:
+    headers = [
+        "问题组排名", "问题组", "记录类型", "相似问法/清洗后问题", "出现次数",
+        "用户", "手机号", "用户ID", "会话ID", "消息ID", "原始问题", "提问时间",
+    ]
+    sheet.append(headers)
+    for item in items:
+        for variant in item.get("variants") or []:
+            sheet.append([
+                item.get("rank"), _excel_safe(item.get("title")), "相似问法",
+                _excel_safe(variant.get("text")), variant.get("count", 0),
+                "", "", "", "", "", "", "",
+            ])
+        for sample in item.get("samples") or []:
+            sheet.append([
+                item.get("rank"), _excel_safe(item.get("title")), "原始样例",
+                _excel_safe(sample.get("cleaned")), "",
+                _excel_safe(sample.get("userName")), _excel_safe(sample.get("userPhone")),
+                sample.get("userId"), sample.get("sessionId"), sample.get("messageId"),
+                _excel_safe(sample.get("raw")), _excel_datetime(sample.get("createdAt")),
+            ])
+    _style_table(
+        sheet,
+        widths=(12, 38, 12, 46, 12, 16, 18, 12, 12, 12, 72, 20),
+        table_name="QuestionInsightSamples",
+    )
+    for row in sheet.iter_rows(min_row=2):
+        for index in (1, 3, 10):
+            row[index].alignment = Alignment(vertical="top", wrap_text=True)
+        row[11].number_format = "yyyy-mm-dd hh:mm:ss"
+
+
+def _append_summary(sheet, result: dict) -> None:
+    range_info = result.get("range") or {}
+    summary = result.get("summary") or {}
+    sheet.append(["项目", "内容"])
+    rows = [
+        ("开始时间", range_info.get("dateFrom") or "未限制"),
+        ("结束时间", range_info.get("dateTo") or "未限制"),
+        ("单次清洗上限", range_info.get("maxMessages", 0)),
+        ("最低频次", summary.get("minCount", 0)),
+        ("纳入清洗消息", summary.get("scannedMessages", 0)),
+        ("有效问题", summary.get("cleanedQuestions", 0)),
+        ("过滤低价值", summary.get("filteredMessages", 0)),
+        ("全部问题组", summary.get("clusterCount", 0)),
+        ("导出问题组", summary.get("visibleClusterCount", 0)),
+        ("清洗规则版本", summary.get("cleanerVersion", "")),
+        ("导出时间", datetime.now()),
+        ("说明", "导出结果按所选日期范围和最低频次生成，包含全部符合条件的问题组，不受页面分页影响。"),
+    ]
+    for label, value in rows:
+        sheet.append([label, _excel_safe(value) if isinstance(value, str) else _excel_datetime(value)])
+    _style_table(sheet, widths=(22, 88), table_name="QuestionInsightSummary")
+    for row in sheet.iter_rows(min_row=2):
+        row[1].alignment = Alignment(vertical="top", wrap_text=True)
+        if isinstance(row[1].value, datetime):
+            row[1].number_format = "yyyy-mm-dd hh:mm:ss"
+
+
+def _style_table(sheet, *, widths: tuple[int, ...], table_name: str) -> None:
+    sheet.freeze_panes = "A2"
+    sheet.sheet_view.showGridLines = False
+    header_fill = PatternFill("solid", fgColor="176B57")
+    for cell in sheet[1]:
+        cell.fill = header_fill
+        cell.font = Font(color="FFFFFF", bold=True)
+        cell.alignment = Alignment(horizontal="center", vertical="center", wrap_text=True)
+    sheet.row_dimensions[1].height = 30
+    for index, width in enumerate(widths, start=1):
+        sheet.column_dimensions[_column_letter(index)].width = width
+    if sheet.max_row > 1:
+        ref = f"A1:{_column_letter(len(widths))}{sheet.max_row}"
+        table = Table(displayName=table_name, ref=ref)
+        table.tableStyleInfo = TableStyleInfo(
+            name="TableStyleMedium4",
+            showRowStripes=True,
+            showFirstColumn=False,
+            showLastColumn=False,
+        )
+        sheet.add_table(table)
+
+
+def _column_letter(index: int) -> str:
+    result = ""
+    while index:
+        index, remainder = divmod(index - 1, 26)
+        result = chr(65 + remainder) + result
+    return result
+
+
+def _excel_safe(value: object | None) -> str:
+    text = str(value or "")
+    return f"'{text}" if text.lstrip().startswith(("=", "+", "-", "@")) else text
+
+
+def _excel_datetime(value: object) -> object:
+    if isinstance(value, datetime) and value.tzinfo is not None:
+        return value.replace(tzinfo=None)
+    return value
--- a/ai_knowledge_base_v2/apps/backend/tests/test_admin_pagination.py
+++ b/ai_knowledge_base_v2/apps/backend/tests/test_admin_pagination.py
@@ -3,9 +3,10 @@ from datetime import datetime, timedelta
 from sqlalchemy.orm import Session
 from sqlalchemy.pool import StaticPool
 from types import SimpleNamespace
+from openpyxl import load_workbook

 from app.api.admin_agent_records import attention_list, retrieval_logs
-from app.api.admin_records import ai_logs, chat_detail, chat_messages, question_insights, refresh_question_insights
+from app.api.admin_records import ai_logs, chat_detail, chat_messages, export_question_insights, question_insights, refresh_question_insights
 from app.api.admin_users import list_users, user_operation_detail, user_topic_options
 from app.models import Base
 from app.models.chat import ChatMessage, ChatSession, TopicSession
@@ -14,6 +15,9 @@ from app.models.insight import QuestionInsightCleanedQuestion
 from app.models.knowledge import HumanAttentionRecord, KnowledgeRetrievalLog
 from app.models.logs import AiRequestLog
 from app.models.user import User
+from app.services.question_insight_export_service import QuestionInsightExportService
+from fastapi import HTTPException
+import pytest


 def _database() -> Session:
@@ -211,3 +215,24 @@ def test_question_insights_clean_and_cluster_similar_user_questions():
        assert data["items"][0]["category"] == "homework"
        assert data["items"][0]["noHitCount"] == 1
        assert data["items"][0]["needsKnowledgeFollowUp"] is True
+
+        workbook = load_workbook(QuestionInsightExportService.build(data), data_only=True)
+        assert workbook.sheetnames == ["洞察结果", "相似问法与样例", "统计说明"]
+        assert workbook["洞察结果"]["B2"].value == data["items"][0]["title"]
+        assert workbook["洞察结果"]["D2"].value == 2
+        assert workbook["洞察结果"]["J2"].value == "是"
+        assert workbook["相似问法与样例"].max_row > 2
+        assert workbook["统计说明"]["B10"].value == 1
+
+
+def test_question_insight_export_rejects_reversed_date_range():
+    with _database() as db, pytest.raises(HTTPException) as exc:
+        export_question_insights(
+            dateFrom=datetime(2026, 8, 2),
+            dateTo=datetime(2026, 8, 1),
+            minCount=2,
+            maxMessages=100,
+            db=db,
+            current_admin=SimpleNamespace(id=1),
+        )
+    assert exc.value.status_code == 400
--- a/ai_knowledge_base_v2/apps/backend/tests/test_admin_permissions.py
+++ b/ai_knowledge_base_v2/apps/backend/tests/test_admin_permissions.py
@@ -4,7 +4,9 @@ import json

 import pytest
 from fastapi import HTTPException
+from starlette.requests import Request

+from app.core.dependencies import enforce_admin_access
 from app.models.admin import Admin, Role
 from app.services.admin_permission_service import ALL_PERMISSION_CODES, permissions_for, require_permission

@@ -24,3 +26,24 @@ def test_role_permissions_are_restricted_to_catalog() -> None:
    with pytest.raises(HTTPException) as exc:
        require_permission(admin, "users.delete")
    assert exc.value.status_code == 403
+
+
+def test_agent_options_only_require_agent_view_permission() -> None:
+    role = Role(code="agent_tester", name="Agent 测试员", permissions=json.dumps(["prompt.view", "prompt.batch"]))
+    admin = Admin(
+        id=3,
+        username="agent-tester",
+        password="hash",
+        name="Agent 测试员",
+        status=1,
+        must_change_password=0,
+        is_super_admin=0,
+        role=role,
+    )
+    request = Request({"type": "http", "method": "GET", "path": "/api/admin/agent/knowledge-options", "headers": []})
+    assert enforce_admin_access(request, admin) is admin
+
+    knowledge_request = Request({"type": "http", "method": "GET", "path": "/api/admin/knowledge/options", "headers": []})
+    with pytest.raises(HTTPException) as exc:
+        enforce_admin_access(knowledge_request, admin)
+    assert exc.value.status_code == 403
--- a/ai_knowledge_base_v2/apps/backend/tests/test_agent_batch_tests.py
+++ b/ai_knowledge_base_v2/apps/backend/tests/test_agent_batch_tests.py
@@ -4,15 +4,17 @@ import json
 import pytest
 from fastapi import HTTPException, UploadFile
 from openpyxl import Workbook, load_workbook
-from sqlalchemy import create_engine, select
+from sqlalchemy import create_engine, func, select
 from sqlalchemy.orm import Session
 from sqlalchemy.pool import StaticPool

+from app.api.admin_agent_batch import _job_or_404, list_batch_jobs
 from app.models import Base
 from app.models.admin import Admin
-from app.models.agent_batch import AgentBatchTestItem
+from app.models.agent_batch import AgentBatchTest, AgentBatchTestItem
 from app.models.ai_config import ModelConfig
 from app.services.agent_batch_test_service import AgentBatchTestService, MAX_BATCH_QUESTIONS
+from app.services.agent_batch_test_worker import AgentBatchTestWorker


 def _database() -> Session:
@@ -98,6 +100,8 @@ def test_create_job_snapshots_config_and_export_keeps_failed_rows() -> None:

        assert json.loads(job.generation_config)["maxToken"] == 8192
        assert json.loads(job.knowledge_names) == ["全部正式开放知识库"]
+        assert job.execution_mode == "sequential"
+        assert job.concurrency_limit == 1

        exported = AgentBatchTestService.export_workbook(job, items)
        workbook = load_workbook(exported, data_only=True)
@@ -123,3 +127,111 @@ def test_import_rejects_excel_formula_question() -> None:

    assert exc.value.status_code == 400
    assert "不能使用 Excel 公式" in exc.value.detail
+
+
+def test_sequential_job_only_claims_one_running_item() -> None:
+    with _database() as db:
+        db.add_all([_admin(), _model()])
+        db.commit()
+        job = _create_test_job(db, execution_mode="sequential", concurrency_limit=1)
+        db.commit()
+
+        first_id = AgentBatchTestWorker.claim_next(db, worker_id="worker-a")
+        assert first_id is not None
+        assert AgentBatchTestWorker.claim_next(db, worker_id="worker-b") is None
+
+        first = db.get(AgentBatchTestItem, first_id)
+        first.status = "success"
+        db.commit()
+        second_id = AgentBatchTestWorker.claim_next(db, worker_id="worker-b")
+        assert second_id is not None
+        assert db.get(AgentBatchTestItem, second_id).batch_id == job.id
+
+
+def test_concurrent_job_respects_per_job_limit() -> None:
+    with _database() as db:
+        db.add_all([_admin(), _model()])
+        db.commit()
+        _create_test_job(db, execution_mode="concurrent", concurrency_limit=2, question_count=3)
+        db.commit()
+
+        first_id = AgentBatchTestWorker.claim_next(db, worker_id="worker-a")
+        second_id = AgentBatchTestWorker.claim_next(db, worker_id="worker-b")
+        assert first_id is not None
+        assert second_id is not None
+        assert first_id != second_id
+        assert AgentBatchTestWorker.claim_next(db, worker_id="worker-c") is None
+        running = db.scalar(
+            select(func.count(AgentBatchTestItem.id))
+            .where(AgentBatchTestItem.status == "running")
+        )
+        assert running == 2
+
+
+def test_jobs_keep_each_admins_knowledge_snapshot_separate() -> None:
+    with _database() as db:
+        first_admin = _admin()
+        second_admin = Admin(id=2, username="reviewer", password="hash", name="评审管理员", status=1)
+        db.add_all([first_admin, second_admin, _model()])
+        db.commit()
+        first = _create_test_job(db, admin=first_admin, knowledge_ids=[])
+        second = _create_test_job(db, admin=second_admin, knowledge_ids=[])
+        db.commit()
+
+        assert first.created_by == 1
+        assert second.created_by == 2
+        assert first.id != second.id
+        assert first.knowledge_ids == "[]"
+        assert second.knowledge_ids == "[]"
+
+
+def test_normal_admin_only_lists_and_opens_own_jobs() -> None:
+    with _database() as db:
+        first_admin = _admin()
+        second_admin = Admin(id=2, username="reviewer", password="hash", name="评审管理员", status=1)
+        super_admin = Admin(id=3, username="root", password="hash", name="超级管理员", status=1, is_super_admin=1)
+        db.add_all([first_admin, second_admin, super_admin, _model()])
+        db.commit()
+        own_job = _create_test_job(db, admin=first_admin)
+        other_job = _create_test_job(db, admin=second_admin)
+        db.commit()
+
+        own_result = list_batch_jobs(page=1, pageSize=10, db=db, current_admin=first_admin)["data"]
+        assert own_result["total"] == 1
+        assert [item["id"] for item in own_result["items"]] == [own_job.id]
+        assert _job_or_404(db, own_job.id, first_admin).id == own_job.id
+        with pytest.raises(HTTPException) as exc:
+            _job_or_404(db, other_job.id, first_admin)
+        assert exc.value.status_code == 404
+
+        super_result = list_batch_jobs(page=1, pageSize=10, db=db, current_admin=super_admin)["data"]
+        assert super_result["total"] == 2
+        assert _job_or_404(db, other_job.id, super_admin).id == other_job.id
+
+
+def _create_test_job(
+    db: Session,
+    *,
+    admin: Admin | None = None,
+    execution_mode: str = "sequential",
+    concurrency_limit: int = 1,
+    question_count: int = 2,
+    knowledge_ids: list[int] | None = None,
+) -> AgentBatchTest:
+    return AgentBatchTestService.create_job(
+        db,
+        admin=admin or db.get(Admin, 1),
+        filename="并发测试.xlsx",
+        rows=[
+            {"rowNumber": index + 2, "externalNo": str(index + 1), "question": f"问题 {index + 1}"}
+            for index in range(question_count)
+        ],
+        config={
+            "name": "并发稳定性测试",
+            "promptContent": "根据知识库回答",
+            "modelId": 1,
+            "knowledgeIds": knowledge_ids or [],
+            "executionMode": execution_mode,
+            "concurrencyLimit": concurrency_limit,
+        },
+    )
--- a/ai_knowledge_base_v2/docker-compose.dev.yml
+++ b/ai_knowledge_base_v2/docker-compose.dev.yml
@@ -60,6 +60,7 @@ services:
      AGENT_BATCH_WORKER_ENABLED: "true"
      AGENT_BATCH_POLL_SECONDS: "2"
      AGENT_BATCH_STALE_MINUTES: "30"
+      AGENT_BATCH_WORKER_CONCURRENCY: "10"
      JWT_SECRET_KEY: local-dev-secret-change-before-production
      MOCK_SMS_ENABLED: "true"
      MOCK_SMS_CODE: "123456"
--- a/ai_knowledge_base_v2/docker-compose.prod.yml
+++ b/ai_knowledge_base_v2/docker-compose.prod.yml
@@ -59,6 +59,7 @@ services:
      AGENT_BATCH_WORKER_ENABLED: ${AGENT_BATCH_WORKER_ENABLED:-true}
      AGENT_BATCH_POLL_SECONDS: ${AGENT_BATCH_POLL_SECONDS:-2}
      AGENT_BATCH_STALE_MINUTES: ${AGENT_BATCH_STALE_MINUTES:-30}
+      AGENT_BATCH_WORKER_CONCURRENCY: ${AGENT_BATCH_WORKER_CONCURRENCY:-10}
      JWT_SECRET_KEY: ${JWT_SECRET_KEY:?必须配置 JWT_SECRET_KEY}
      CONFIG_ENCRYPTION_KEY: ${CONFIG_ENCRYPTION_KEY:?必须配置 CONFIG_ENCRYPTION_KEY}
      MOCK_SMS_ENABLED: "false"