From a208dfabe95717c46dd9d61a5eb4cdc8df5487ee Mon Sep 17 00:00:00 2001 From: eezd Date: Thu, 5 Nov 2026 23:09:42 +0800 Subject: [PATCH] Fix long-input evidence extraction and trim assignment context --- .../src/alphapulse/news_research/pipeline.py | 108 +++++++++++++----- .../src/alphapulse/news_research/service.py | 7 +- backend/tests/test_news_pipeline.py | 92 ++++++++++++--- backend/tests/test_news_research_workspace.py | 12 +- backend/tests/test_time_repair.py | 8 +- scripts/smoke_news_research.py | 7 +- 6 files changed, 177 insertions(+), 57 deletions(-) diff --git a/backend/src/alphapulse/news_research/pipeline.py b/backend/src/alphapulse/news_research/pipeline.py index 674411d..bda8145 100644 --- a/backend/src/alphapulse/news_research/pipeline.py +++ b/backend/src/alphapulse/news_research/pipeline.py @@ -1,6 +1,7 @@ """Evidence contracts and lossless context preparation for event research.""" import json -from typing import Literal +import re +from typing import Annotated, Literal from pydantic import BaseModel, Field, ConfigDict from alphapulse.news_research.time_contracts import PackageTimePoint from alphapulse.news_research.model import complete, fits_context, SubmissionFailure @@ -68,55 +69,110 @@ def validate_event(result, articles): if claim.verdict != 'insufficient' and not claim.evidence_refs: raise ValueError('核验结论缺少原文证据') -async def extract_articles(provider, system, articles): - pending = [(a['id'], a['title']+'\n'+(a.get('summary') or '')+'\n'+(a.get('content') or '')) for a in articles] +class EvidenceSelection(BaseModel): + model_config = ConfigDict(extra='forbid') + facts: str = Field(min_length=1) + evidence_ids: list[Annotated[int, Field(ge=0, strict=True)]] = Field(min_length=1) + + +SELECTION_RULES = '\n本次只输出中文facts和evidence_ids，选择本次输入提供的证据编号；不生成quote或article_id。程序将按编号回填逐字原文，不得选择无关证据。' + + +def selected_extraction(selection, evidence): + ids = selection.evidence_ids + if len(ids) != len(set(ids)) or not set(ids) <= evidence.keys(): + raise ValueError('提取结果含未知或重复证据编号') + return Extraction(facts=selection.facts, evidence_refs=[evidence[i] for i in ids]) + + +def source_passages(article): + # Separate source fields so no quote can span a synthesized title/body boundary. + # Sentence boundaries keep evidence readable; long sentences still retain every character. + passages = [] + for field in ('title', 'summary', 'content'): + for sentence in re.split(r'(?<=[。！？])|(?<=[.!?])(?=[ \t])', article.get(field) or ''): + passages.extend({'field': field, 'text': sentence[start:start+500]} + for start in range(0, len(sentence), 500)) + return passages + + +async def extract_articles(provider, system, articles, *, scopes=None): + pending = [(article, source_passages(article)) for article in articles] extracts, usages = [], [] + rules = system + SELECTION_RULES try: - schema = Extraction while pending: - identifier, content = pending.pop(0) - part = {'task': 'extract', 'article_id': identifier, 'content': content} - if not fits_context(provider, system, part, schema): - if len(content) < 2: - raise ValueError('模型上下文不足以读取分段') - middle = len(content)//2 - pending[0:0] = [(identifier, content[:middle]), (identifier, content[middle:])] + article, passages = pending.pop(0) + identifier = article['id'] + if not passages: + raise ValueError('文章缺少可读取原文') + part = {'task': 'extract', 'article_id': identifier, + 'packages': [p for p in (scopes or []) if p['news_id'] == identifier], + 'passages': [{'id': i, **passage} for i, passage in enumerate(passages)]} + if not fits_context(provider, rules, part, EvidenceSelection): + if len(passages) > 1: + middle = len(passages)//2 + halves = passages[:middle], passages[middle:] + elif passages and len(passages[0]['text']) > 1: + passage = passages[0] + middle = len(passage['text'])//2 + halves = ([{**passage, 'text': passage['text'][:middle]}], + [{**passage, 'text': passage['text'][middle:]}]) + else: + raise ValueError(f'文章{identifier}的归属边界及最小原文片段超出模型上下文') + pending[0:0] = [(article, half) for half in halves] continue - extract, used = await complete(provider, system, part, schema) + selection, used = await complete(provider, rules, part, EvidenceSelection) usages.append(used) - validate_evidence(extract.evidence_refs, {a['id']: a for a in articles}) - if any(r.article_id != identifier or r.quote not in content for r in extract.evidence_refs): - raise ValueError('分段证据越界') - if extract.evidence_refs: - extracts.append(extract.model_dump()) + evidence = {i: Evidence(article_id=identifier, quote=p['text']) for i, p in enumerate(passages)} + extract = selected_extraction(selection, evidence) + validate_evidence(extract.evidence_refs, {identifier: article}) + extracts.append(extract.model_dump()) return extracts, usages except Exception as error: - raise SubmissionFailure(str(error), {'parts': usages, 'failed': getattr(error, 'usage', None)}) from error + raise SubmissionFailure(f'文章{identifier}长输入原文提取失败：{error}', {'parts': usages, 'failed': getattr(error, 'usage', None)}) from error + + +def consolidation_input(extracts): + evidence, identities, parts = {}, {}, [] + for extract in extracts: + ids = [] + for ref in extract['evidence_refs']: + key = (ref['article_id'], ref['quote']) + if key not in identities: + identifier = len(evidence) + identities[key] = identifier + evidence[identifier] = Evidence.model_validate(ref) + ids.append(identities[key]) + parts.append({'facts': extract['facts'], 'evidence_ids': ids}) + return {'task': 'consolidate', 'extractions': parts, + 'evidence': [{'id': i, **ref.model_dump()} for i, ref in evidence.items()]}, evidence + async def compact_extractions(provider, system, payload, schema, fits=fits_context): usages = [] try: - rules = system + '\nconsolidate任务只合并给定分段事实，保留冲突和时间线；每个article_id至少保留一条给定原文引用，不新增引用，压缩重复表述。' + rules = system + '\nconsolidate任务只合并给定分段事实，保留冲突和时间线；每个article_id至少保留一条给定原文证据，压缩重复表述。' + SELECTION_RULES while not fits(provider, system, payload, schema): extracts = payload['extractions'] groups, group = [], [] for extract in extracts: - if not fits_context(provider, rules, {'task': 'consolidate', 'extractions': [*group, extract]}, Extraction): + if not fits_context(provider, rules, consolidation_input([*group, extract])[0], EvidenceSelection): if not group: raise ValueError('模型上下文不足以归纳分段结果，请增加上下文配置') groups.append(group); group = [] - if not fits_context(provider, rules, {'task': 'consolidate', 'extractions': [extract]}, Extraction): + if not fits_context(provider, rules, consolidation_input([extract])[0], EvidenceSelection): raise ValueError('模型上下文不足以归纳分段结果，请增加上下文配置') group.append(extract) if group: groups.append(group) compacted = [] for group in groups: - result, used = await complete(provider, rules, {'task': 'consolidate', 'extractions': group}, Extraction) + request, evidence = consolidation_input(group) + selection, used = await complete(provider, rules, request, EvidenceSelection) usages.append(used) - refs = {(r['article_id'], r['quote']) for part in group for r in part['evidence_refs']} - actual = {(r.article_id, r.quote) for r in result.evidence_refs} - if not actual <= refs or {r[0] for r in actual} != {r[0] for r in refs}: - raise ValueError('分层归纳遗漏来源或改写原文引用') + result = selected_extraction(selection, evidence) + if {ref.article_id for ref in result.evidence_refs} != {ref.article_id for ref in evidence.values()}: + raise ValueError('分层归纳遗漏来源') compacted.append(result.model_dump()) if len(json.dumps(compacted, ensure_ascii=False).encode()) >= len(json.dumps(extracts, ensure_ascii=False).encode()): raise ValueError('分层归纳未缩减输入，无法在模型上下文内完整汇总') diff --git a/backend/src/alphapulse/news_research/service.py b/backend/src/alphapulse/news_research/service.py index 78e5f1d..fb6e18f 100644 --- a/backend/src/alphapulse/news_research/service.py +++ b/backend/src/alphapulse/news_research/service.py @@ -373,7 +373,6 @@ async def candidate_context(session, catalog): "id": e.id, "revision": e.revision, "title": e.title, - "tags": e.tags, "reporting_range": None, "evidence": [], } @@ -876,12 +875,12 @@ async def process_event(identifier, config, force=False): if not researcher.fits(provider, skill, request, ResearchResult): extraction_rules = ( skill - + "\n当前事件及各成员归属边界：" - + json.dumps({"event": payload["event"], "packages": payload["packages"]}, ensure_ascii=False) + + "\n当前事件：" + + json.dumps(payload["event"], ensure_ascii=False) + "\nextract只提取当前片段对本事件有用的事实与连续原文引用，不作评级。consolidate保留各来源、冲突、时间与引用，不新增事实。" ) extracts, preparations = await extract_articles( - provider, extraction_rules, payload["articles"] + provider, extraction_rules, payload["articles"], scopes=payload["packages"] ) if { ref["article_id"] diff --git a/backend/tests/test_news_pipeline.py b/backend/tests/test_news_pipeline.py index 40051a4..839f585 100644 --- a/backend/tests/test_news_pipeline.py +++ b/backend/tests/test_news_pipeline.py @@ -23,7 +23,7 @@ def event_result(pipeline, refs): 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']) +@pytest.mark.parametrize('invalid', ['missing_source', 'unknown_reference', 'no_progress']) def test_compaction_rejects_lost_provenance_and_stops_without_progress(pipeline, monkeypatch, invalid): provider = SimpleNamespace(max_tokens=100, context_window=8000) extracts = [{'facts': 'x'*3000, 'evidence_refs': [{'article_id': i, 'quote': 'original'}]} for i in (1, 2)] @@ -31,14 +31,14 @@ def test_compaction_rejects_lost_provenance_and_stops_without_progress(pipeline, async def complete(provider, system, payload, schema): calls.append(payload) - refs = [r for e in payload['extractions'] for r in e['evidence_refs']] - if invalid == 'missing_source': refs = refs[:1] - if invalid == 'rewritten_quote': refs[0] = {**refs[0], 'quote': 'invented'} + ids = [r['id'] for r in payload['evidence']] + if invalid == 'missing_source': ids = ids[:1] + if invalid == 'unknown_reference': ids[0] = 999 facts = 'x'*10000 if invalid == 'no_progress' else 'compressed' - return pipeline.Extraction(facts=facts, evidence_refs=refs), {} + return schema(facts=facts, evidence_ids=ids), {} monkeypatch.setattr(pipeline, 'complete', complete) - message = '未缩减输入' if invalid == 'no_progress' else '遗漏来源或改写' + message = {'no_progress': '未缩减输入', 'missing_source': '遗漏来源', 'unknown_reference': '未知或重复证据编号'}[invalid] with pytest.raises(ValueError, match=message): asyncio.run(pipeline.compact_extractions(provider, 'rules', {'task': 'event', 'id': 1, 'extractions': extracts}, pipeline.ReviewResult)) assert len(calls) == 1 @@ -57,20 +57,80 @@ def test_extraction_requires_original_evidence(pipeline, monkeypatch): 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: []} + articles = [{'id': i, 'title': f'来源{i}', 'summary': '原文摘要。', 'content': '完整原文。' * 5000} for i in (1, 2)] + fragments = {i: {field: [] for field in ('title', 'summary', 'content')} for i in (1, 2)} + scopes = [{'id': 10+i, 'news_id': i, 'scope_quote': f'来源{i}', 'other_scopes': []} for i in (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} + identifier = payload['article_id'] + assert payload['packages'] == [scopes[identifier-1]] + for passage in payload['passages']: + fragments[identifier][passage['field']].append(passage['text']) + assert all(len(p['text']) <= 500 for p in payload['passages']) + return schema(facts='保留事实。', evidence_ids=[0]), {'input_tokens': 10} monkeypatch.setattr(pipeline, 'complete', complete) - extracts, usage = asyncio.run(pipeline.extract_articles(provider, '按原文提取事实。', articles)) + extracts, usage = asyncio.run(pipeline.extract_articles(provider, '按原文提取事实。', articles, scopes=scopes)) 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()) + for field in ('title', 'summary', 'content'): + assert ''.join(fragments[article['id']][field]) == article[field] + assert len(fragments[article['id']]['content']) > 1 + assert len(usage) == len(extracts) assert {ref['article_id'] for part in extracts for ref in part['evidence_refs']} == {1, 2} + pipeline.validate_evidence([pipeline.Evidence.model_validate(ref) for part in extracts for ref in part['evidence_refs']], {a['id']: a for a in articles}) + + +@pytest.mark.parametrize('selection', [[999], [0, 0]]) +def test_extraction_rejects_unknown_or_repeated_ids_and_retains_usage(pipeline, monkeypatch, selection): + async def complete(provider, rules, payload, schema): + return schema(facts='事实。', evidence_ids=selection), {'input_tokens': 17, 'output_tokens': 3} + monkeypatch.setattr(pipeline, 'complete', complete) + with pytest.raises(pipeline.SubmissionFailure, match='未知或重复证据编号') as failure: + asyncio.run(pipeline.extract_articles(SimpleNamespace(max_tokens=100, context_window=12000), '规则', + [{'id': 42, 'title': '原文标题', 'summary': '原文摘要', 'content': '原文正文'}])) + assert failure.value.usage['parts'] == [{'input_tokens': 17, 'output_tokens': 3}] + + +def test_selected_quotes_preserve_fields_whitespace_and_foreign_language(pipeline, monkeypatch): + article = {'id': 42, 'title': 'Official title', 'summary': 'Another field', 'content': 'Binance\n(SignalPlus) access was cut.\nNext report.'} + async def complete(provider, rules, payload, schema): + assert schema is pipeline.EvidenceSelection + assert '不生成quote或article_id' in rules + return schema(facts='币安切断访问。', evidence_ids=[p['id'] for p in payload['passages']]), {} + monkeypatch.setattr(pipeline, 'complete', complete) + extracts, _ = asyncio.run(pipeline.extract_articles(SimpleNamespace(max_tokens=100, context_window=12000), '规则', [article])) + quotes = [r['quote'] for part in extracts for r in part['evidence_refs']] + assert ''.join(quotes) == article['title']+article['summary']+article['content'] + assert all(any(quote in article[field] for field in ('title','summary','content')) for quote in quotes) + with pytest.raises(ValueError, match='原文引用'): + pipeline.validate_evidence([pipeline.Evidence(article_id=42, quote='Official title\nAnother field')], {42:article}) + + +def test_extraction_schema_never_accepts_model_generated_quotes(pipeline): + with pytest.raises(ValueError): + pipeline.EvidenceSelection.model_validate({'facts':'事实。','evidence_ids':[0], + 'evidence_refs':[{'article_id':42,'quote':'编造引文'}]}) + + +def test_single_long_sentence_splits_without_losing_characters(pipeline, monkeypatch): + provider = SimpleNamespace(max_tokens=100, context_window=1500) + article = {'id': 9, 'title': '', 'summary': '', 'content': '长' * 500} + received = [] + async def complete(provider, rules, payload, schema): + assert pipeline.fits_context(provider, rules, payload, schema) + received.extend(p['text'] for p in payload['passages']) + return schema(facts='分段事实。', evidence_ids=[0]), {} + monkeypatch.setattr(pipeline, 'complete', complete) + asyncio.run(pipeline.extract_articles(provider, '规则', [article])) + assert len(received) > 1 and ''.join(received) == article['content'] + + +def test_oversized_scope_fails_before_model_call_with_article_identity(pipeline, monkeypatch): + from unittest.mock import AsyncMock + complete = AsyncMock() + monkeypatch.setattr(pipeline, 'complete', complete) + with pytest.raises(pipeline.SubmissionFailure, match='文章9.*最小原文片段超出模型上下文'): + asyncio.run(pipeline.extract_articles(SimpleNamespace(max_tokens=100, context_window=1500), '规则', + [{'id': 9, 'title': '标题', 'content': '正文'}], scopes=[{'news_id': 9, 'scope_quote': '长' * 5000}])) + complete.assert_not_awaited() diff --git a/backend/tests/test_news_research_workspace.py b/backend/tests/test_news_research_workspace.py index 4d6081c..6d99661 100644 --- a/backend/tests/test_news_research_workspace.py +++ b/backend/tests/test_news_research_workspace.py @@ -2,6 +2,7 @@ from datetime import UTC, datetime, timedelta from types import SimpleNamespace as N +import json import pytest @@ -648,13 +649,12 @@ def test_long_event_compaction_uses_final_research_context(trading, monkeypatch, monkeypatch.setattr(service, "validate_timeline", lambda *args: None) async def consolidate(provider, rules, payload, schema): - assert schema is pipeline.Extraction - assert "当前事件及各成员归属边界" in rules and quote in rules + assert schema is pipeline.EvidenceSelection + assert "当前事件：" in rules and "当前事件及各成员归属边界" not in rules assert "extract和consolidate生成的facts等叙述也必须用简体中文" in rules assert "time_text及取自原文的period_text逐字保留原语言和字符" in rules assert pipeline.fits_context(provider, rules, payload, schema) - refs = [ref for part in payload["extractions"] for ref in part["evidence_refs"]] - return pipeline.Extraction(facts="调查继续。", evidence_refs=refs), {"mock": True} + return schema(facts="调查继续。", evidence_ids=[ref["id"] for ref in payload["evidence"]]), {"mock": True} complete = AsyncMock(side_effect=consolidate) monkeypatch.setattr(pipeline, "complete", complete) @@ -679,7 +679,9 @@ def test_long_event_compaction_uses_final_research_context(trading, monkeypatch, compressed = {**versions[0].input, "articles": [], "extractions": extracts} reader = service.Research() assert not reader.fits(provider, skill, versions[0].input, ResearchResult) - assert not reader.fits(provider, boundary_rules, compressed, ResearchResult) + assert len(extract.await_args.kwargs["scopes"]) == 40 + assert all(p["keywords"] == keywords for p in extract.await_args.kwargs["scopes"]) + assert len(boundary_rules) < len(json.dumps(versions[0].input["packages"], ensure_ascii=False)) assert reader.fits(provider, skill, compressed, ResearchResult) != needs_compaction diff --git a/backend/tests/test_time_repair.py b/backend/tests/test_time_repair.py index 5d570d9..d631c65 100644 --- a/backend/tests/test_time_repair.py +++ b/backend/tests/test_time_repair.py @@ -101,7 +101,7 @@ def usage_tokens(usage): def test_extraction_failure_keeps_completed_and_failed_call_usage(trading, monkeypatch, failure): from unittest.mock import AsyncMock from alphapulse.news_research import pipeline, model as service - extract = pipeline.Extraction(facts='Source', evidence_refs=[pipeline.Evidence(article_id=1, quote='Source')]) + extract = pipeline.EvidenceSelection(facts='Source', evidence_ids=[0 if failure == 'second_call' else 999]) complete = AsyncMock(side_effect=[(extract, {'input_tokens': 10}), service.SubmissionFailure('synthetic failure', {'input_tokens': 40})]) monkeypatch.setattr(pipeline, 'complete', complete) monkeypatch.setattr(pipeline, 'fits_context', lambda *args: True) @@ -109,6 +109,7 @@ def test_extraction_failure_keeps_completed_and_failed_call_usage(trading, monke with pytest.raises(service.SubmissionFailure) as error: asyncio.run(pipeline.extract_articles(SimpleNamespace(), 'rules', articles)) assert usage_tokens(error.value.usage) == (50 if failure == 'second_call' else 10) + assert complete.await_count == (2 if failure == 'second_call' else 1) @pytest.mark.parametrize('failure', ['second_call', 'evidence', 'no_shrink']) @@ -116,14 +117,15 @@ def test_compaction_failure_keeps_every_paid_call(trading, monkeypatch, failure) from unittest.mock import AsyncMock from alphapulse.news_research import pipeline, model as service group = {'facts': 'Very long source fact ' * 50, 'evidence_refs': [{'article_id': 1, 'quote': 'Source'}]} - extract = pipeline.Extraction(facts='Source' if failure != 'no_shrink' else 'Longer' * 1000, - evidence_refs=[pipeline.Evidence(article_id=1, quote='Changed' if failure == 'evidence' else 'Source')]) + extract = pipeline.EvidenceSelection(facts='Source' if failure != 'no_shrink' else 'Longer' * 1000, + evidence_ids=[999 if failure == 'evidence' else 0]) complete = AsyncMock(side_effect=[(extract, {'input_tokens': 10}), service.SubmissionFailure('synthetic failure', {'input_tokens': 40})]) monkeypatch.setattr(pipeline, 'complete', complete) monkeypatch.setattr(pipeline, 'fits_context', lambda *args: True) with pytest.raises(service.SubmissionFailure) as error: asyncio.run(pipeline.compact_extractions(SimpleNamespace(), 'rules', {'extractions': [group]}, pipeline.ReviewResult, fits=lambda *args: False)) assert usage_tokens(error.value.usage) == (50 if failure == 'second_call' else 10) + assert complete.await_count == (2 if failure == 'second_call' else 1) diff --git a/scripts/smoke_news_research.py b/scripts/smoke_news_research.py index b89d927..3861b0b 100644 --- a/scripts/smoke_news_research.py +++ b/scripts/smoke_news_research.py @@ -579,6 +579,7 @@ async def main(): await s.commit() event = await s.get(Event, event_id) context = (await service.candidate_context(s, [event]))[0] + assert "tags" not in context and context["id"] == event.id and context["title"] == event.title assert context["reporting_range"] == { "first": (now - timedelta(days=180)).isoformat(), "latest": now.isoformat(), @@ -766,9 +767,9 @@ async def review_regressions(config, complete, package_id, news_id): ).output # Force compression, then exercise both member read tools with frozen originals. - async def extract(provider, rules, articles): - assert "当前事件及各成员归属边界" in rules - assert '"keyword_canonicals": ["serialincident"]' in rules + async def extract(provider, rules, articles, *, scopes): + assert "当前事件：" in rules + assert all(p["keyword_canonicals"] == ["serialincident"] for p in scopes) return [ { "evidence_refs": [