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

scripts/evaluate-knowledge-quality.test.mjs

@@ -1,13 +1,15 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
 import {
-  evaluateCase, ndcgAtK, parseArgs, recallAtK, reciprocalRank, summarize, thresholdFailures, validateDataset
+  evaluateCase, ndcgAtK, normalizeSourceName, parseArgs, recallAtK, reciprocalRank, summarize,
+  thresholdFailures, validateDataset
 } from './evaluate-knowledge-quality.mjs';
 
 test('ranking metrics use fragment ids and preserve top-k order', () => {
   assert.equal(recallAtK(['a', 'b', 'c'], ['b', 'd']), 0.5);
   assert.equal(reciprocalRank(['a', 'b'], ['b']), 0.5);
   assert.ok(ndcgAtK(['b', 'a'], ['b', 'c']) > 0.61);
+  assert.equal(ndcgAtK(['b', 'b', 'b'], ['b']), 1);
 });
 
 test('forbidden fragments and sources are surfaced as leakage', () => {
@@ -25,7 +27,7 @@ test('forbidden fragments and sources are surfaced as leakage', () => {
 test('no-answer cases measure false positives separately from answerable recall', () => {
   const answerable = evaluateCase([{ fragmentId: 1, sourceName: 'SOP' }], {
     id: 'answerable', relevantFragmentIds: [1], requiredSourceNames: ['SOP']
-  }, 5);
+  }, 5, false);
   const noAnswer = evaluateCase([{ fragmentId: 9, sourceName: 'Unrelated' }], {
     id: 'no-answer', expectedNoEvidence: true, relevantFragmentIds: [], caseType: 'no-answer'
   }, 5);
@@ -35,6 +37,45 @@ test('no-answer cases measure false positives separately from answerable recall'
   assert.equal(summary.noAnswerPrecision, 0);
 });
 
+test('candidate recall is independent from the final no-evidence gate and display prefixes', () => {
+  const result = evaluateCase([
+    {
+      fragmentId: 122126,
+      sourceName: '第 3 段：银城物业费用报销发票管理操作手册.pdf',
+      sourceAuthority: 'COMPANY_POLICY',
+      sourceKind: 'OPERATING_MANUAL',
+      selectedForEvidence: false
+    }
+  ], {
+    id: 'finance',
+    requiredSourceNames: ['银城物业费用报销发票管理操作手册'],
+    requiredSourceAuthorities: ['COMPANY_POLICY'],
+    requiredSourceKinds: ['OPERATING_MANUAL'],
+    expectedNoEvidence: false
+  }, 20, true);
+  assert.equal(result.recallAt5, 1);
+  assert.equal(result.noEvidence, true);
+  assert.equal(normalizeSourceName('第 12 段：制度.PDF'), '制度');
+});
+
+test('same-name ungoverned copies do not satisfy formal source recall', () => {
+  const result = evaluateCase([
+    {
+      fragmentId: 102169,
+      sourceName: '银城物业费用报销发票管理操作手册.pdf',
+      sourceAuthority: 'UNKNOWN',
+      sourceKind: 'REFERENCE_MATERIAL'
+    }
+  ], {
+    id: 'finance-authority',
+    requiredSourceNames: ['银城物业费用报销发票管理操作手册'],
+    requiredSourceAuthorities: ['COMPANY_POLICY'],
+    requiredSourceKinds: ['OPERATING_MANUAL']
+  }, 20, true);
+  assert.equal(result.recallAt20, 0);
+  assert.equal(result.wrongSource, true);
+});
+
 test('threshold failures can fail a release gate', () => {
   const failures = thresholdFailures({
     recallAtK: 0.7, mrr: 0.5, ndcgAtK: 0.6, forbiddenLeakageRate: 0.1,
@@ -49,6 +90,7 @@ test('threshold failures can fail a release gate', () => {
 test('dataset contract distinguishes answerable and no-answer cases', () => {
   assert.doesNotThrow(() => validateDataset({ schemaVersion: 1, cases: [
     { id: 'a', query: 'q', relevantFragmentIds: [1] },
+    { id: 'source', query: 'q-source', requiredSourceNames: ['正式手册'] },
     { id: 'b', query: 'q2', relevantFragmentIds: [], expectedNoEvidence: true }
   ] }));
   assert.throws(() => validateDataset({ schemaVersion: 1, cases: [