feat(aihr): harden finance RAG retrieval and evidence gates

2026-08-03 22:34:18 +08:00
parent 99e8c509d5
commit 721839c847
22 changed files with 1306 additions and 153 deletions
@@ -0,0 +1,64 @@
+{
+  "schemaVersion": 1,
+  "datasetCode": "finance-reimbursement-rag-regression-v1",
+  "cases": [
+    {
+      "id": "finance-500-colloquial",
+      "query": "我有个报销500块钱，告诉我怎么走流程。",
+      "caseType": "formal-finance-ambiguous-basis",
+      "requiredSourceNames": ["银城物业费用报销发票管理操作手册"],
+      "requiredSourceAuthorities": ["COMPANY_POLICY"],
+      "requiredSourceKinds": ["OPERATING_MANUAL"],
+      "forbiddenSourceNames": ["200块私了", "500块维修"],
+      "expectedNoEvidence": true
+    },
+    {
+      "id": "finance-500-small-purchase",
+      "query": "500元以下的零星采购怎么报销？",
+      "caseType": "formal-finance-ambiguous-basis",
+      "requiredSourceNames": ["银城物业费用报销发票管理操作手册"],
+      "requiredSourceAuthorities": ["COMPANY_POLICY"],
+      "requiredSourceKinds": ["OPERATING_MANUAL"],
+      "forbiddenSourceNames": ["200块私了", "500块维修"],
+      "expectedNoEvidence": true
+    },
+    {
+      "id": "finance-500-chinese-amount",
+      "query": "五百块的小额自采怎么走账？",
+      "caseType": "formal-finance-ambiguous-basis",
+      "requiredSourceNames": ["银城物业费用报销发票管理操作手册"],
+      "requiredSourceAuthorities": ["COMPANY_POLICY"],
+      "requiredSourceKinds": ["OPERATING_MANUAL"],
+      "forbiddenSourceNames": ["200块私了", "500块维修"],
+      "expectedNoEvidence": true
+    },
+    {
+      "id": "finance-private-settlement-no-answer",
+      "query": "200块维修费用能不能直接私了后报销？",
+      "caseType": "no-answer",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["200块私了", "500块维修"]
+    },
+    {
+      "id": "finance-over-limit-unsupported",
+      "query": "零星采购超过500元怎么报销？",
+      "caseType": "unsupported-threshold",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["200块私了", "500块维修"]
+    },
+    {
+      "id": "finance-monthly-basis-conflict",
+      "query": "月累计500元以内的自采都能报销吗？",
+      "caseType": "amount-basis-conflict",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["员工访谈", "200块私了", "500块维修"]
+    },
+    {
+      "id": "finance-per-transaction-basis-conflict",
+      "query": "单笔500元以内的自采都能报销吗？",
+      "caseType": "amount-basis-conflict",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["员工访谈", "200块私了", "500块维修"]
+    }
+  ]
+}