eezd/AlphaPulse
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor(news): remove retired research workflow
Development CI / check-and-build (push) Successful in 2m53s
Details
GitHub Git mirror / mirror (push) Successful in 7s
Details
Development CI / deploy-development (push) Successful in 2m31s
Details

Browse Source
This commit is contained in:
eezd committed 2026-11-05 18:38:25 +08:00
1 parent f3894c3728
commit d79a10c2e6
69 files changed
+460 -6713
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

backend/tests/test_news_pipeline.py
+26 -190
View File
@@ -1,45 +1,15 @@
"""Two-stage contracts reject ambiguous grouping and ungrounded conclusions."""
import asyncio
from types import SimpleNamespace
from unittest.mock import AsyncMock
import pytest
@pytest.fixture
def pipeline(trading):
from alphapulse.news_risk import pipeline
from alphapulse.news_research import pipeline
return pipeline
def match(pipeline, **changes):
return dict(event=dict(event_id=None,group_key='event',title='公告',fact='公告',reason='同一具体事项'),package_id=None, group_key='event', title='公告', category='macro', entities=['央行'],
fact='公告', evidence='公告', urgent=False, tags=['央行'], content_kind='report', **changes)
@pytest.mark.parametrize('package_id,group_key', [(None, None), (1, 'event')])
def test_grouping_target_is_exclusive(pipeline, package_id, group_key):
data = match(pipeline)
data.update(package_id=package_id, group_key=group_key)
with pytest.raises(ValueError): pipeline.Match(**data)
def test_all_domains_require_nonempty_distinct_packages(pipeline):
item = pipeline.Match(**match(pipeline))
with pytest.raises(ValueError, match='重复关联'):
pipeline.Grouping(reason='政治公告', packages=[item, item])
with pytest.raises(ValueError):
pipeline.Grouping(reason='说明', packages=[])
assert pipeline.Grouping(reason='文化说明', packages=[item]).packages
@pytest.mark.parametrize('scope,enabled,published', [('shared', True, 1), ('news', False, 1), ('news', True, None)])
def test_unpublished_disabled_and_shared_skills_never_reach_model(pipeline, scope, enabled, published):
node = SimpleNamespace(scope=scope, kind='skill', published_version_id=published, enabled=enabled, parent_id=None)
session = SimpleNamespace(get=AsyncMock(return_value=node))
with pytest.raises(ValueError):
asyncio.run(pipeline.configuration(session, SimpleNamespace(grouping_skill_id=1), 'grouping'))
def test_evidence_cannot_cross_article_boundaries(pipeline):
articles = {1: {'title': '央行维持利率'}, 2: {'title': '交易所暂停提现'}}
with pytest.raises(ValueError, match='原文引用'):
@@ -47,45 +17,10 @@ def test_evidence_cannot_cross_article_boundaries(pipeline):
pipeline.validate_evidence([pipeline.Evidence(article_id=2, quote='暂停提现')], articles)
def test_old_backfill_is_not_selected_by_ingestion_time(pipeline):
from datetime import UTC, datetime
from sqlalchemy import select
from sqlalchemy.dialects.postgresql import dialect
from alphapulse.models import NewsItem
from alphapulse.news_risk.worker import eligible
sql = str(select(NewsItem.id).where(eligible(datetime.now(UTC))).compile(dialect=dialect()))
assert 'greatest(news_items.published_at, news_items.source_updated_at)' in sql
assert 'first_seen_at' not in sql and 'last_seen_at' not in sql
def event_result(pipeline, refs):
return pipeline.EventResult(id=1, title='Synthetic event', summary='Synthetic summary', category='macro', entities=[],
return pipeline.ReviewResult(id=1, title='Synthetic event', summary='Synthetic summary', category='macro', entities=[],
risk_stars=1, importance_stars=4, direction='unknown', verification='insufficient', confidence='low', reasoning='Synthetic', uncertainty='Synthetic',
timeline_points=[dict(occurred_at=None, time_precision='unknown', time_text='Unknown', reported_at=None, description='Synthetic', verification='insufficient', evidence_refs=[pipeline.Evidence.model_validate(r).model_dump() for r in refs])], facts='Synthetic', timeline='Synthetic', asset_impacts='Unknown', opinions='Unknown', counter_evidence='Unknown', watchpoints='Unknown', claims=[pipeline.Claim(statement='Synthetic', verdict='insufficient', evidence_refs=[])], related_packages=[], excluded_article_ids=[], evidence_refs=refs)
def test_long_event_uses_bounded_layers_and_retains_all_sources(pipeline, monkeypatch):
provider = SimpleNamespace(max_tokens=2000, context_window=12000)
articles = [{'id': i+1, 'title': 'Synthetic', 'summary': '', 'content': 'b'*20000} for i in range(20)]
calls = []
async def complete(provider, system, payload, schema):
assert pipeline.fits_context(provider, system, payload, schema)
calls.append(payload['task'])
if payload['task'] == 'extract':
return pipeline.Extraction(facts='x'*600, evidence_refs=[pipeline.Evidence(
article_id=payload['article_id'], quote=payload['content'].strip()[0])]), {}
refs = {r['article_id']: r for e in payload['extractions'] for r in e['evidence_refs']}
if payload['task'] == 'consolidate':
return pipeline.Extraction(facts='Consolidated facts and conflicts', evidence_refs=list(refs.values())), {}
assert set(refs) == {a['id'] for a in articles}
return event_result(pipeline, list(refs.values())), {}
monkeypatch.setattr(pipeline, 'complete', complete)
result, usage = asyncio.run(pipeline.summarize_articles(provider, 'rules', articles, {'id': 1, 'fact': 'Synthetic event'}))
assert result.id == 1 and calls[-1] == 'event'
assert calls.count('extract') > len(articles) and calls.count('consolidate') > 1
assert len(usage['parts']) == len(calls)-1
timeline_points=[dict(occurred_at=None, time_precision='unknown', time_text='Unknown', reported_at=None, description='Synthetic', verification='insufficient', evidence_refs=[pipeline.Evidence.model_validate(r).model_dump() for r in refs])], facts='Synthetic', timeline='Synthetic', asset_impacts='Unknown', opinions='Unknown', counter_evidence='Unknown', watchpoints='Unknown', claims=[pipeline.Claim(statement='Synthetic', verdict='insufficient', evidence_refs=[])], excluded_article_ids=[], evidence_refs=refs)
@pytest.mark.parametrize('invalid', ['missing_source', 'rewritten_quote', 'no_progress'])
@@ -105,131 +40,11 @@ def test_compaction_rejects_lost_provenance_and_stops_without_progress(pipeline,
monkeypatch.setattr(pipeline, 'complete', complete)
message = '未缩减输入' if invalid == 'no_progress' else '遗漏来源或改写'
with pytest.raises(ValueError, match=message):
asyncio.run(pipeline.compact_extractions(provider, 'rules', {'task': 'event', 'id': 1, 'extractions': extracts}, pipeline.EventResult))
asyncio.run(pipeline.compact_extractions(provider, 'rules', {'task': 'event', 'id': 1, 'extractions': extracts}, pipeline.ReviewResult))
assert len(calls) == 1
def test_large_candidate_catalog_does_not_extract_or_drop_short_article(pipeline, monkeypatch):
provider = SimpleNamespace(max_tokens=2000, context_window=12000)
article = {'id': 1, 'title': 'Synthetic', 'summary': '', 'content': 'Original text'}
candidates = [{'id': i+1, 'title': 'x'*200, 'published_at': '2026-10-31T00:00:00Z', 'entities': ['entity']} for i in range(80)]
complete = AsyncMock(side_effect=AssertionError('Candidate overflow must not trigger source extraction'))
monkeypatch.setattr(pipeline, 'complete', complete)
payload, usage = asyncio.run(pipeline.grouping_payload(provider, 'rules', article, candidates))
assert payload['article'] == article and not usage
assert 0 < len(payload['candidates']) < len(candidates)
assert payload['candidates'] == candidates[:len(payload['candidates'])]
assert pipeline.fits_context(provider, 'rules', payload, pipeline.Grouping)
corrected = pipeline.pack_candidates(provider, 'rules', {**payload, 'correction': 'Invalid output '*100}, payload['candidates'])
assert pipeline.fits_context(provider, 'rules', corrected, pipeline.Grouping)
assert len(corrected['candidates']) < len(payload['candidates'])
complete.assert_not_awaited()
def test_long_grouping_source_is_compacted_before_candidates(pipeline, monkeypatch):
provider = SimpleNamespace(max_tokens=2000, context_window=12000)
article = {'id': 1, 'title': 'Synthetic', 'summary': '', 'content': 'b'*200000}
calls = []
async def complete(provider, system, payload, schema):
assert pipeline.fits_context(provider, system, payload, schema)
calls.append(payload['task'])
if payload['task'] == 'extract':
return pipeline.Extraction(facts='x'*600, evidence_refs=[pipeline.Evidence(article_id=1, quote='b')]), {}
return pipeline.Extraction(facts='Compressed facts', evidence_refs=[pipeline.Evidence(article_id=1, quote='b')]), {}
monkeypatch.setattr(pipeline, 'complete', complete)
payload, _ = asyncio.run(pipeline.grouping_payload(provider, 'rules', article, [{'id': 2, 'title': 'Candidate'}]))
assert 'consolidate' in calls
assert payload['article_id'] == 1 and payload['candidates'][0]['id'] == 2
assert pipeline.fits_context(provider, 'rules', payload, pipeline.Grouping)
@pytest.mark.parametrize('length', [50, 20000])
def test_package_scope_survives_direct_and_extracted_roundup(pipeline, monkeypatch, length):
provider = SimpleNamespace(max_tokens=2000, context_window=12000)
article = {'id': 1, 'title': 'Roundup', 'summary': '', 'content': 'b'*length}
observed = []
async def complete(provider, system, payload, schema):
observed.append(payload)
if payload['task'] == 'extract':
return schema(facts=payload['package']['fact'], evidence_refs=[pipeline.Evidence(article_id=1, quote='b')]), {}
return event_result(pipeline, [pipeline.Evidence(article_id=1, quote='b')]).model_copy(update={'id': payload['id']}), {}
monkeypatch.setattr(pipeline, 'complete', complete)
for identifier, fact in [(1, '央行公布利率'), (2, '公司发布新产品')]:
package = {'id': identifier, 'title': fact, 'fact': fact, 'tags': [fact], 'category': 'general', 'content_kind': 'report'}
before = len(observed)
result, _ = asyncio.run(pipeline.summarize_articles(provider, 'rules', [article], package))
assert result.id == identifier
assert all(payload['package'] == package for payload in observed[before:])
assert observed[-1]['package']['fact'] == '公司发布新产品'
def test_long_event_reserves_real_research_tools_before_extraction(pipeline, monkeypatch):
from alphapulse.news_risk import research
provider = SimpleNamespace(max_tokens=1000, context_window=12000, reasoning_effort=None)
article = {'id': 1, 'title': 'Report', 'summary': '', 'content': ''}
package = {'id': 1, 'fact': '单个事实'}
while pipeline.fits_context(provider, 'rules', {'task': 'event', 'id': 1, 'package': package,
'articles': [{**article, 'content': article['content']+'b'*100}]}, pipeline.EventResult):
article['content'] += 'b'*100
review = research.Research()
assert not review.fits(provider, 'rules', {'task': 'event', 'id': 1, 'package': package, 'articles': [article]}, pipeline.EventResult)
extracted = []
async def complete(provider, system, payload, schema):
extracted.append(payload)
return pipeline.Extraction(facts='原文中的事实', evidence_refs=[pipeline.Evidence(article_id=1, quote='b')]), {}
async def stream(provider, system, messages, tools, effort):
assert review.fits_messages(provider, system, messages, pipeline.EventResult)
assert 'extractions' in messages[0]['text']
result = event_result(pipeline, [pipeline.Evidence(article_id=1, quote='b')])
yield research.llm.Completion(tool_calls=[{'id': 'done', 'name': 'submit', 'arguments': result.model_dump()}])
monkeypatch.setattr(pipeline, 'complete', complete)
monkeypatch.setattr(research.llm, 'stream', stream)
result, _ = asyncio.run(pipeline.summarize_articles(provider, 'rules', [article], package, review))
assert result.id == 1 and extracted
@pytest.mark.parametrize('has_related_fact', [True, False])
def test_package_roundup_skips_empty_slices_but_requires_related_evidence(pipeline, monkeypatch, has_related_fact):
provider = SimpleNamespace(max_tokens=1000, context_window=18000)
article = {'id': 1, 'title': 'Roundup', 'summary': '',
'content': ('RATE decision. ' if has_related_fact else '')+'Unrelated product announcement. '*1500}
package = {'id': 1, 'fact': 'RATE decision'}
system = pipeline.BASE+'\n'+pipeline.TEMPLATES['analysis'][1]
extracted, submitted = [], []
async def complete(provider, system, payload, schema):
if payload['task'] == 'extract':
related = 'RATE' in payload['content']
extracted.append(related)
return schema.model_validate({'facts': 'RATE decision' if related else '',
'evidence_refs': [{'article_id': 1, 'quote': 'RATE'}] if related else []}), {'input_tokens': 1}
assert payload['task'] == 'event'
submitted.append(payload)
assert all(e['facts'] and e['evidence_refs'] for e in payload['extractions'])
return event_result(pipeline, [pipeline.Evidence(article_id=1, quote='RATE')]), {}
monkeypatch.setattr(pipeline, 'complete', complete)
if has_related_fact:
result, usage = asyncio.run(pipeline.summarize_articles(provider, system, [article], package))
assert result.evidence_refs[0].quote == 'RATE' and len(submitted) == 1
assert len(usage['parts']) == len(extracted)
assert len(submitted[0]['extractions']) == sum(extracted)
else:
with pytest.raises(ValueError, match='未提取到当前包相关证据'):
asyncio.run(pipeline.summarize_articles(provider, system, [article], package))
assert not submitted
assert len(extracted) > 1 and False in extracted
@pytest.mark.parametrize('facts,refs', [('Unsupported fact', []), ('', [{'article_id': 1, 'quote': 'Original'}])])
def test_package_extraction_requires_facts_and_evidence_together(pipeline, facts, refs):
with pytest.raises(ValueError, match='分段事实与证据必须同时为空或同时提供'):
pipeline.PackageExtraction(facts=facts, evidence_refs=refs)
def test_grouping_extraction_and_final_summary_still_require_evidence(pipeline, monkeypatch):
def test_extraction_requires_original_evidence(pipeline, monkeypatch):
async def complete(provider, system, payload, schema):
return schema.model_validate({'facts': '', 'evidence_refs': []}), {}
monkeypatch.setattr(pipeline, 'complete', complete)
@@ -238,3 +53,24 @@ def test_grouping_extraction_and_final_summary_still_require_evidence(pipeline,
[{'id': 1, 'title': 'Original', 'content': '', 'summary': ''}]))
with pytest.raises(ValueError):
event_result(pipeline, [])
def test_long_extraction_preserves_complete_originals_and_source_ids(pipeline, monkeypatch):
provider = SimpleNamespace(max_tokens=1000, context_window=12000)
articles = [{'id': i, 'title': f'来源{i}', 'summary': '', 'content': '完整原文。' * 5000} for i in (1, 2)]
fragments = {1: [], 2: []}
async def complete(provider, rules, payload, schema):
assert pipeline.fits_context(provider, rules, payload, schema)
identifier, content = payload['article_id'], payload['content']
fragments[identifier].append(content)
return pipeline.Extraction(facts='保留事实。', evidence_refs=[
pipeline.Evidence(article_id=identifier, quote='完整原文。')]), {'input_tokens': 10}
monkeypatch.setattr(pipeline, 'complete', complete)
extracts, usage = asyncio.run(pipeline.extract_articles(provider, '按原文提取事实。', articles))
for article in articles:
assert ''.join(fragments[article['id']]) == article['title']+'\n\n'+article['content']
assert len(fragments[article['id']]) > 1
assert len(usage) == len(extracts) == sum(len(parts) for parts in fragments.values())
assert {ref['article_id'] for part in extracts for ref in part['evidence_refs']} == {1, 2}