优化:未分类连续维护与柏拉图预汇总,同步已发布修复

This commit is contained in:
2026-11-01 12:40:50 +08:00
parent 572873c6b8
commit 38290f5fec
135 changed files with 17613 additions and 663 deletions
@@ -0,0 +1,99 @@
"""No-draft formatting preserves real Markdown structure and scoped rank counts."""
import copy
import json
import sys
import threading
import unittest
from pathlib import Path
sys.path[:0] = [str(Path(__file__).resolve().parents[1] / 'scripts'), str(Path(__file__).parent)]
from nao_native_answer import NativeCompositionError, _validate_blocks
from nao_evidence_finalizer import synthesize_native
from test_agent_answer import evidence
from test_native_content_title import response
class NoDraftStructureTests(unittest.TestCase):
def test_short_lead_with_later_markdown_survives_full_composition_and_summary(self):
for detail in ['\n'.join(['- 各组织分别保留正式期间、业务范围和重点表格。'] * 18),
'\n\n'.join(['各组织分别保留正式期间、业务范围和重点表格。'] * 18)]:
calls = []
text = '综合合格率95.00%。\n\n' + detail
def provider(request, timeout):
body = json.loads(request.data); calls.append(body)
if body['messages'][0]['content'].startswith('逐句核验证据'):
return response({'reviews': [{'supported': True, 'statisticalClaimsSupported': True}],
'missingGoalDescriptions': []})
return response({'blocks': [{'role': 'core', 'text': text, 'factIds': ['F1']}]})
result = synthesize_native(base_url='https://offline.invalid/v1', api_key='fixture', model='fixture',
question='合格率分析', evidence=evidence(), cancel_event=threading.Event(), opener=provider,
context={'_agentTask': {'kind': 'analysis', 'executionMode': 'native_agent'}})
self.assertTrue(result['evidenceSufficient'], result['compositionDiagnostics'])
self.assertEqual(len(calls), 2)
self.assertFalse(result['compositionDiagnostics']['formatRepairAttempted'])
self.assertIn(detail, result['answer'])
def test_lead_limit_checks_the_first_paragraph_and_keeps_later_list(self):
facts = [{'id': 'F1', 'text': '已确认的正式结果。', 'kind': 'overall',
'label': '正式结果', 'evidenceRef': 'Q00000001'}]
for separator in ['\n\n', '\n']:
text = '各范围的报废表现应分别解读。' + separator + '\n'.join(
['- 保留各自期间、组织范围和正式口径,重点表格紧随对应结论。'] * 15)
with self.subTest(separator=separator):
blocks, selected = _validate_blocks({'blocks': [{'role': 'core', 'text': text,
'factIds': ['F1']}]}, facts, {})
self.assertEqual(blocks[0]['text'], text)
self.assertEqual(selected, ['F1'])
with self.assertRaisesRegex(NativeCompositionError, '首段过长'):
_validate_blocks({'blocks': [{'role': 'core', 'text': '结论' * 170,
'factIds': ['F1']}]}, facts, {})
def test_top_count_uses_the_verified_group_without_combining_other_scopes(self):
facts = [{'id': 'F' + str(i), 'text': '已确认规格表现。', 'kind': 'product',
'label': '规格' + chr(64 + i), 'evidenceRef': 'history-row-' + str(i),
'rank': i, 'rankGroupKey': 'source-a/section-0'} for i in range(1, 6)]
def validate(count, selected=facts):
return _validate_blocks({'blocks': [{'role': 'core', 'text': '保留产品Top' + str(count) + '。',
'factIds': [fact['id'] for fact in selected]}]}, selected, {})
self.assertEqual(validate(5)[0][0]['text'], '保留产品Top5。')
with self.assertRaises(NativeCompositionError): validate(6)
mismatched = copy.deepcopy(facts)
mismatched[-1]['rankGroupKey'] = 'source-b/section-0'
with self.assertRaises(NativeCompositionError): validate(5, mismatched)
other = [{**fact, 'id': 'F' + str(index + 6), 'rankGroupKey': 'source-b/section-0',
'evidenceRef': 'other-history-row-' + str(index)} for index, fact in enumerate(facts)]
full_pool = facts + other
with self.assertRaises(NativeCompositionError):
_validate_blocks({'blocks': [{'role': 'core', 'text': '本次产品Top5依次如下。',
'factIds': ['F1', 'F2', 'F8', 'F9', 'F10']}]}, full_pool, {})
def test_partial_scope_title_joins_bounded_metadata_repair_and_keeps_verified_body(self):
calls, generation = [], []
def provider(request, timeout):
body = json.loads(request.data); calls.append(body)
if body['messages'][0]['content'].startswith('逐句核验证据'):
return response({'reviews': [{'supported': True, 'statisticalClaimsSupported': True}],
'missingGoalDescriptions': []})
generation.append(body)
guide = json.loads(body['messages'][1]['content'])['reportTitleGuide']
self.assertEqual(guide['excludedScopeLabels'], ['产业部甲'])
if len(generation) == 2:
self.assertIn('titleMetadataOnlyRepair', body['messages'][-1]['content'])
return response({'reportTitle': '2026年8—9月合格率变化'})
return response({'reportTitle': '产业部甲质量变化', 'blocks': [{'role': 'core',
'text': '综合合格率95.00%。', 'factIds': ['F1']}]})
result = synthesize_native(base_url='https://offline.invalid/v1', api_key='fixture', model='fixture',
question='汇总合格率变化', evidence=evidence(), cancel_event=threading.Event(), opener=provider,
context={'_agentTask': {'kind': 'report', 'executionMode': 'native_agent'},
'_historyTitleExcludedScopeLiterals': ['产业部甲']})
self.assertEqual(len(calls), 3)
self.assertEqual(len(generation), 2)
self.assertTrue(result['bodyEvidenceSufficient'])
self.assertTrue(result['titleValid'])
self.assertTrue(result['evidenceSufficient'])
self.assertEqual(result['reportTitle'], '2026年8—9月合格率变化')
self.assertIn('95.00%', result['answer'])
if __name__ == '__main__':
unittest.main()