Files
quality_analyse/runtime/dsh/application/quality_native_run.py
T
lyun bc19475b29 修复:DSH 自主查询、图片直显与报告稳定性
允许 Agent 在循环中查目录、修正查询和提问,保留实时权限检查及模型原始回答。
统一相同查询证据与图表资源;排名补齐阶段指标,修正 OR 重量筛选,并限制模型明细输入。
增加运行恢复、图表自适应坐标轴与两位小数,修复报告主图选择、重复图及边缘标签。

验证:独立副本通过 162 项 Python、73 项前端及 21 项后端测试,前端构建通过。
线上已用原 8 题复测,并追加两轮双期对比;TOP100 与组合筛选分别核对 100 行和 20 行。
2026-11-12 16:14:20 +08:00

264 lines
19 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""Publish DSH's original answer without content review or recomposition."""
from __future__ import annotations
import copy
import json
import time
import re
def initial_read_arguments(task, question):
"""Read the confirmed business frame before asking a model for prose."""
if task.get('agentDriven') or task.get('kind') == 'report' or task.get('freshQuery') is not True or not task.get('fields'):
return None
dimensions = task.get('dimensions') or []
if task.get('aggregateSelection') and not task.get('comparisonPeriod') and not task.get('comparisonRequired'):
return {'mode': 'query', 'view': task['aggregateSelection']['dimension']}
stages = task.get('stages') or []
count = re.search(r'(?:top|前)\s*(\d+)', question, re.I)
weights = [field for field in task.get('fields', []) if field in ('totalWeight', 'scrapWeight')]
if (count and not task.get('comparisonPeriod') and not task.get('comparisonRequired')
and len(dimensions) == 1 and dimensions[0] in ('department', 'group', 'product')
and len(stages) == 1 and len(weights) == 1):
return {'mode': 'rank', 'dimension': dimensions[0], 'stage': stages[0],
'metric': weights[0], 'limit': min(100, max(1, int(count[1])))}
return {'mode': 'compare' if task.get('comparisonPeriod') else 'query'}
def _documents(directory, excluded):
"""Read this attempt's saved tool results for optional chart rendering."""
documents = []
for path in directory.glob('quality-evidence-*.json'):
if path.name in excluded or path.is_symlink():
continue
try:
document = json.loads(path.read_text())
if isinstance(document, dict) and isinstance(document.get('result'), dict):
documents.append(document)
except (OSError, ValueError):
continue
return documents
def image_resources(visuals):
"""工具只提供由正式数据实际构造的本轮图表资源。"""
return [{'src': 'chart:' + visual['id'], 'alt': visual['title'],
'scope': copy.deepcopy(visual.get('scope', {})),
'period': copy.deepcopy(visual.get('period', {}))}
for visual in visuals.get('visuals', [])
if isinstance(visual, dict) and isinstance(visual.get('id'), str)
and isinstance(visual.get('title'), str)]
def run_native_exploration(*, request, task, authority, actor_root, delegated, connection,
cancel_event, emit, prior_evidence_files, phase_timings=None):
from dsh_agent_runtime import run_dsh_agent
from quality_agent_task import bind_task
from quality_native_exploration import NativeExplorationTools, native_model_view, _proofs
from quality_protocol_execution import _identity
from quality_runtime_context import _atomic_json, _read_json, ConversationAccumulator
from quality_runtime_policy import turn_deadline
from quality_tools import tool_definition, QualityToolError
timings = dict(phase_timings or {})
identity = _identity(authority)
context = bind_task({key: copy.deepcopy(value) for key, value in request['context'].items()
if key in ('_serverHistory', '_serverQuestions', '_protocol')}, task)
context.update(releaseId=request['releaseId'], _nativeExecutionId=identity['executionId'])
history = task.get('kind') == 'report'
history_snapshot = {}
if history:
started = time.monotonic()
history_snapshot = delegated.call('/api/analysis/protocol/history', payload={}, method='POST')
timings['historySnapshot'] = round((time.monotonic() - started) * 1000)
context['historySnapshot'] = history_snapshot
context['readPolicy'] = '本轮仅使用服务器提供的历史快照生成报告,不重新查询业务数据。'
request['context'] = context
context['analysisDirectory'] = copy.deepcopy(request.get('_analysisDirectory',{}))
_atomic_json(actor_root / 'private/last-agent-task.json', {
'sessionId': request['sessionId'], 'releaseId': request['releaseId'],
'task': task, 'protocol': identity})
tools = None if history else NativeExplorationTools(
delegated=delegated, evidence_dir=actor_root / 'evidence', task=task,
identity={**identity, 'releaseId': request['releaseId']},
state_path=actor_root / 'private/native-exploration-state.json', directory=context['analysisDirectory'])
definition = tool_definition(native_protocol=True, task=task, release_id=request['releaseId'])
allowed_modes = definition['inputSchema']['properties']['mode']['enum']
context['toolPolicy'] = {'allowedModes': allowed_modes,
'guidance': '总体query一次已含全部阶段率、逐月趋势、投入、报废和正式大类。query不传stage,period及comparisonPeriod可省略,由运行时带入;若传必须用对象,不用JSON字符串。需要产业部、业务分组时分别query(view=department/group)一次批量汇总。排名用rank且metric必填:投入重量totalWeight,报废重量scrapWeight。产业部排名用dimension=department,不要先选某个产业部。rank返回行含投入、报废、合格率等辅助字段,不为同一对象再查另一个指标排名或扩大Top数量。按原问题取得足够数据即回答,不为未要求的组织或产品扩展调查。analysisDirectory.periods是全部已发布月份,不等于本轮窗口。双期比较用compare,并明确period和comparisonPeriod;一次多月query已提供各月trend时也可直接对比,不需重复读取。窗口合计不是焦点月份值。重复查询返回reusedEvidence时直接使用已有结果,不重复工具循环。'}
if task.get('agentDriven'):
context['agentIntent'] = copy.deepcopy(request.get('_agentIntent', {}))
context['toolPolicy']['guidance'] += (' 本轮发布版本固定,相同查询复用已返回数据和图片,freshQuery不能产生更新数据。'
'图表src是已经就绪的展示资源,写入正文即可,不需要再次查询来获取或确认图片。'
'queryCompletion描述本次调用的完成状态;数据已覆盖问题时直接回答,只有需要新的范围、期间、指标或分页时再查询。')
context['toolPolicy']['guidance'] += (' 本轮由你自主确定和修正查询。agentIntent仅为未确认的意图提示,不是取数前提。'
'规格缩写先catalog(dimension=product,search=原规格)查询授权候选;只有一个候选时使用完整name,有多个候选时直接向用户提问,保留原目标与期间,不猜选。'
'需要不同期间、筛选条件或指标时可在下一次工具调用中指定;参数错误和未匹配名称是可恢复反馈,按返回说明调整后继续。'
'query可传fields=[rate/scrapRate/totalWeight/scrapWeight]表达需要呈现的指标;所有正式值仍来自服务器。只有forbidden才停止查询。')
if task.get('aggregateSelection'):
context['aggregateSelection'] = copy.deepcopy(task['aggregateSelection'])
context['toolPolicy']['guidance'] += ' 本轮按aggregateSelection查询完整分组,initialObservation.rows已先按汇总重量筛选、再按正式合格率排序并取所需Top N(最多100)。保持返回名次和条件,直接撰写回答;不再用重量rank代替合格率排名,不再扩大读取或把summary当作Top N合计。原始需求见aggregateSelection.requestQuestion,当前消息可能只是补充口径。'
from dsh_model_input import adjacent_month_period
adjacent = adjacent_month_period(request['question'], context)
if adjacent:
context['focusPeriod'] = adjacent['focusPeriod']
context['toolPolicy']['guidance'] += ' 本题以focusPeriod月份为主,用query返回的focusMonth获取当月指标和正式报废大类,用trend作前后对比;不把summary窗口合计当作当月。未明确要求具体来源缺陷时,当月正式大类已能说明报废线索,无需再查窗口defects或组织/产品榜单。总体低点不代表各阶段的合格率或报废率也都最低;每个指标独立按trendSummary判断,不推断两阶段一定同向。'
from quality_report_deliverable import requested
context['responsePolicy'] = ('直接回答用户本轮问题,保留必要的口径说明;普通问答不套报告标题、导读、章节或重复结论,不主动生成报告文件。'
if not requested(request['question'], task) else
'报告按问题自主组织:排名重在名次与集中度,诊断重在焦点月份与相邻月差异,结构专题重在类别占比与缺陷。导读控制在三句以内,正文不重复整段导读;一份明细表不在多个章节重复。只呈现题目需要的内容,不为凑章节扩大查询。')
def authorize():
if cancel_event.is_set():
raise RuntimeError('cancelled')
delegated.call('/api/analysis/model-authorization', method='GET')
def query(arguments):
if history:
raise QualityToolError('本轮仅整理已有历史快照,不允许重新查询业务数据。')
if not isinstance(arguments, dict) or arguments.get('mode') not in allowed_modes:
raise QualityToolError('本轮可用mode为 '+', '.join(allowed_modes)+';请选择可用模式。排名用rank并传metric=totalWeight或scrapWeight。')
arguments = copy.deepcopy(arguments)
page = None
if arguments['mode'] == 'query':
offset = arguments.pop('offset', 0)
if type(offset) is not int or offset < 0:
raise QualityToolError('query.offset必须为非负整数。', code='invalid_arguments')
requested_limit = arguments.get('limit', 100)
if type(requested_limit) is not int or requested_limit < 1:
raise QualityToolError('query.limit必须为正整数。', code='invalid_arguments')
page = {'offset': offset, 'limit': min(100, requested_limit)}
projection_task = {**task, **({'modelPage': page} if page else {})}
observation = native_model_view(tools.quality_data(arguments), compact=True, task=projection_task)
if isinstance(observation, dict) and not observation.get('error') and arguments['mode'] != 'catalog':
reused = observation.get('reusedEvidence') is True and not (page and page['offset'])
observation['queryCompletion'] = {
'state': 'no_data' if observation.get('status') == 'no_data' else 'data_ready',
'reused': reused,
'newEvidence': not reused,
'guidance': ('相同查询已完成,本次复用原数据与图表,没有新增证据。'
if reused else '本次查询已完成,数据可直接用于回答。') +
'若已覆盖用户问题,请现在作答;缺失月份按来源覆盖说明,不为同一条件重查。需要新增范围、期间、指标或分页时仍可继续调用。'}
if observation.get('modelPage', {}).get('hasMore'):
observation['queryCompletion']['guidance'] += ' 本次模型展示有分页;确需其余分组时沿用相同条件并传offset=modelPage.nextOffset。不能用当前页声称全量排名;rank和selectionDisplay的Top结果已在全量范围排序,无需补查。'
try:
from analysis_visual_evidence import protocol_visual_result
documents = _documents(actor_root / 'evidence', prior_evidence_files)
snapshot = delegated.call('/api/analysis/protocol/queries', payload={}, method='POST')
images = image_resources(protocol_visual_result(documents, snapshot.get('scopes', []),
request['question'], request['turnId'], release_id=request['releaseId'], base_context=context))
observation = {**observation, 'presentation': {'images': images,
'guidance': '图片已就绪,无需再查询生成或确认。可用![标题](src)将最贴题的一张图直接放入正文;只引用此列表中的src,不自行编造ID或地址,不重复插入相同图表。没有资源时直接写文字和必要的表格。'}}
except Exception as error:
_atomic_json(actor_root / 'private/chart-resource-diagnostics.json', {'errorType': type(error).__name__})
return observation
started = time.monotonic()
initial = None
arguments = None if history else initial_read_arguments(task, request['question'])
if arguments:
initial = ConversationAccumulator()
call_id = 'initial-' + request['turnId']
initial.feed({'type': 'tool-input-available', 'toolName': 'mcp_call', 'toolCallId': call_id,
'input': {'tool': 'quality_data', 'arguments': arguments}})
emit({'type': 'progress', 'message': '正在读取本题正式数据…',
'toolCalls': initial.tool_calls, 'toolCount': len(initial.tool_calls), 'timeline': initial.timeline})
try:
observation = query(arguments)
except Exception as error:
observation = {'error': {'code': getattr(error, 'code', 'unavailable'),
'message': str(error) if isinstance(error, QualityToolError) else '本题正式数据读取未完成。'}}
initial.feed({'type': 'tool-output-available', 'toolName': 'mcp_call', 'toolCallId': call_id,
'output': observation})
context['initialObservation'] = observation
context['initialObservationCall'] = arguments
emit({'type': 'progress', 'message': '正在根据查询结果生成回答…',
'toolCalls': initial.tool_calls, 'toolCount': len(initial.tool_calls), 'timeline': initial.timeline})
timings['initialRead'] = round((time.monotonic() - started) * 1000)
if initial.tool_calls[0].get('status') == 'completed':
timings['firstUsefulMs'] = sum(timings.get(k, 0) for k in
('planningAndReview', 'taskConfirmation', 'historySnapshot', 'initialRead'))
def progress(event):
if ('firstUsefulMs' not in timings and any(
call.get('status') == 'completed' and (call.get('input') or {}).get('mode') not in (None, 'catalog')
for call in event.get('toolCalls', []))):
timings['firstUsefulMs'] = (sum(timings.get(k, 0) for k in
('planningAndReview', 'taskConfirmation', 'historySnapshot'))
+ round((time.monotonic() - started) * 1000))
emit(event)
emit({'type': 'progress', 'message': '正在生成回答…', 'executionMode': 'native_agent'})
accumulator, status, reason, stopped = run_dsh_agent(
actor_root=actor_root, connection=connection, question=request['question'],
context=context, task=task, identity=identity, turn_id=request['turnId'],
cancel_event=cancel_event, emit=progress, tool_handler=query, authorize=authorize,
tool_definition=definition, deadline=turn_deadline(), initial_accumulator=initial)
timings['investigation'] = round((time.monotonic() - started) * 1000)
answer = accumulator.answer if isinstance(accumulator.answer, str) else ''
if cancel_event.is_set():
status, reason = 'CANCELLED', 'user_cancel'
elif status not in ('COMPLETED', 'PARTIAL', 'FAILED', 'CANCELLED'):
status, reason = 'FAILED', reason or 'runtime_error'
business_calls = [call for call in accumulator.tool_calls
if (call.get('input') or {}).get('mode') not in (None, 'catalog')]
if (not task.get('agentDriven') and status == 'COMPLETED' and business_calls
and not any(call.get('status') == 'completed' for call in business_calls)):
status, reason = 'FAILED', 'tool_queries_failed'
# Charts are optional presentation. A chart/provenance failure cannot
# suppress the model body or change the actual agent execution status.
documents, receipts = [], []
visuals = {'visualSchemaVersion': 1, 'visualState': 'not_requested', 'visuals': []}
if history:
visuals.update({key: copy.deepcopy(history_snapshot[key]) for key in
('visualSchemaVersion', 'visualState', 'visuals') if key in history_snapshot})
elif status != 'CANCELLED':
started = time.monotonic()
try:
from analysis_visual_evidence import protocol_visual_result
documents = _documents(actor_root / 'evidence', prior_evidence_files)
receipts = [proof for document in documents for proof in _proofs(document['result'])]
snapshot = delegated.call('/api/analysis/protocol/queries', payload={}, method='POST')
visuals = protocol_visual_result(documents, snapshot.get('scopes', []),
request['question'], request['turnId'], release_id=request['releaseId'], base_context=context)
if visuals.get('missingEvidence') or visuals.get('visualState') in ('failed', 'unavailable'):
_atomic_json(actor_root / 'private/chart-render-diagnostics.json', {
'state': visuals.get('visualState'), 'missing': visuals.get('missingEvidence', [])})
except Exception as error:
_atomic_json(actor_root / 'private/chart-render-error.json', {'errorType': type(error).__name__})
timings['visuals'] = round((time.monotonic() - started) * 1000)
state_path = actor_root / 'private/native-exploration-state.json'
try:
timings['formalQuery'] = _read_json(state_path).get('queryElapsedMs', 0) if state_path.is_file() else 0
except (OSError, ValueError):
timings['formalQuery'] = 0
timings['synthesis'] = 0
result = {
'type': 'result', 'engine': 'DSH', 'status': status, 'answer': answer,
'agentTask': task, 'protocol': identity, 'executionMode': 'native_agent',
'completionMethod': 'dsh_direct_answer',
'stopReason': 'stop' if status == 'COMPLETED' else str(reason or status).lower(),
'finishReason': (accumulator.finish_reasons[-1] if accumulator.finish_reasons else
'stop' if status == 'COMPLETED' else str(reason or status).lower()),
'nativeStopReason': reason, 'nativeStopSucceeded': stopped,
'actualModel': connection[2], 'phaseTimingsMs': timings,
'toolCalls': accumulator.tool_calls, 'toolCount': len(accumulator.tool_calls),
'timeline': accumulator.timeline, 'protocolEvidence': receipts,
'queryEvidenceRefs': [doc['evidenceRef'] for doc in documents if isinstance(doc.get('evidenceRef'), str)],
'queryEvidenceBindings': [{'evidenceRef': doc['evidenceRef'],
'queryIds': list(dict.fromkeys(p['queryId'] for p in _proofs(doc['result'])))}
for doc in documents if isinstance(doc.get('evidenceRef'), str)],
'agentContext': {**{key: copy.deepcopy(task[key]) for key in
('filters', 'period', 'comparisonPeriod') if task.get(key) is not None},
'releaseId': request['releaseId']},
'visualSchemaVersion': visuals.get('visualSchemaVersion', 1),
'visualState': visuals.get('visualState', 'not_requested'), 'visuals': visuals.get('visuals', []),
}
if history:
for key in ('historySources', 'historySections', 'historyCoverage', 'historyEvidence', 'historyVisualSources'):
if key in history_snapshot:
result[key] = copy.deepcopy(history_snapshot[key])
emit(result)
return result