Files
quality_analyse/scripts/dsh_row_selection.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

159 lines
11 KiB
Python

"""Post-aggregation selections over unchanged, complete formal metric rows."""
from __future__ import annotations
import copy
import math
import operator
import re
from decimal import Decimal
def validate_selection(value, *, pending=False):
if not isinstance(value, dict) or set(value) - {'dimension', 'metric', 'rateStage', 'direction', 'limit', 'weightFilter', 'requestQuestion'}:
raise ValueError('汇总排名条件格式无效')
result = copy.deepcopy(value)
if (result.get('dimension') not in ('product', 'department', 'group')
or result.get('metric') != 'rate' or result.get('rateStage') not in (None, 'black', 'bright')
or result.get('direction') not in ('asc', 'desc')
or type(result.get('limit')) is not int or not 1 <= result['limit'] <= (10000 if pending else 100)):
raise ValueError('汇总合格率排名支持 Top1—100,需明确对象和排序方向')
weight = result.get('weightFilter')
if weight is not None:
if (not isinstance(weight, dict) or set(weight) != {'metric', 'stage', 'operator', 'value'}
or weight.get('metric') not in ('totalWeight', 'scrapWeight')
or weight.get('stage') not in ('black', 'bright', 'any', 'all', None)
or weight.get('stage') is None and not pending
or weight.get('operator') not in ('gt', 'gte', 'lt', 'lte')
or type(weight.get('value')) not in (int, float, Decimal)
or not math.isfinite(weight['value']) or weight['value'] < 0):
raise ValueError('汇总重量条件需明确阶段、比较符及非负 kg 数值')
if isinstance(weight['value'], Decimal):
weight['value'] = float(weight['value'])
if not isinstance(result.get('requestQuestion'), str) or not 1 <= len(result['requestQuestion']) <= 4000:
raise ValueError('汇总排名必须保留原问题')
return result
def selection_hint(question, context=None):
"""Bind literal Top/threshold syntax; let the model phrase any clarification."""
context = context or {}
previous = context.get('previousTask') or context.get('_agentTask') or (context.get('_protocol') or {}).get('previousTask') or {}
if (previous.get('needsClarification') and isinstance(previous.get('aggregateSelection'), dict)
and not re.search(r'新任务|换个问题|另一个问题|不查了|取消(?:分析|任务)', question)):
result = copy.deepcopy(previous['aggregateSelection'])
# A clarification answer may name either stage or an explicit Boolean relation.
black, bright = '黑片' in question, '亮片' in question
if result.get('weightFilter'):
any_stage = bool(re.search(r'任一|任意(?:一|1)|或|(?<![a-z])or(?![a-z])', question, re.I))
all_stages = bool(re.search(r'都|均|同时|且|(?<![a-z])and(?![a-z])|全部', question, re.I))
if any_stage:
result['weightFilter']['stage'] = 'any'
elif all_stages and (black and bright or re.search(r'两阶段|两个阶段', question)):
result['weightFilter']['stage'] = 'all'
elif black != bright and not ('合格率' in question and not re.search(r'重量|投入|满足', question)):
result['weightFilter']['stage'] = 'black' if black else 'bright'
changed_count = re.search(r'(?:top|前)\s*(\d{1,4})', question, re.I)
if changed_count:
result['limit'] = int(changed_count[1])
changed_rate = re.search(r'(黑片|亮片)合格率', question)
if changed_rate:
result['rateStage'] = {'黑片': 'black', '亮片': 'bright'}[changed_rate[1]]
return result
match = re.search(r'(?:top|前)\s*(\d{1,4})', question, re.I)
if not match or not re.search(r'合格率.{0,12}(?:最低|最高)|(?:最低|最高).{0,12}合格率', question):
return None
dimension = next((key for label, key in (('产品', 'product'), ('产业部', 'department'), ('部门', 'department'), ('分组', 'group')) if label in question), None)
if not dimension:
return None
rate = re.search(r'(黑片|亮片)(?:综合)?合格率', question)
result = {'dimension': dimension, 'metric': 'rate', 'rateStage': {'黑片': 'black', '亮片': 'bright'}.get(rate[1]) if rate else None,
'direction': 'asc' if '最低' in question else 'desc', 'limit': int(match[1]), 'requestQuestion': question}
weight = re.search(r'(?P<scope>(?:(?:黑片|亮片)[^。;,,]{0,12})?)(?P<kind>投入重量|生产重量|报废重量|重量)\s*(?P<relation>任一|任意一|其中一个|至少一个|均|都|同时)?\s*(?P<op>大于等于|不低于|至少|小于等于|不超过|大于|超过|小于|低于|>=|<=|>|<|≥|≤)\s*(?P<value>\d+(?:\.\d+)?)\s*(?:kg|公斤|千克)', question, re.I)
if weight:
scope = weight['scope'] + (weight['relation'] or ''); black, bright = '黑片' in scope, '亮片' in scope
stage = ('all' if re.search(r'都|均|同时|且|(?<![a-z])and(?![a-z])', scope, re.I) else 'any' if re.search(r'任一|或|(?<![a-z])or(?![a-z])', scope, re.I) else None) if black and bright else 'black' if black else 'bright' if bright else None
op = weight['op']; op = 'gte' if op in ('大于等于', '不低于', '至少', '>=', '≥') else 'lte' if op in ('小于等于', '不超过', '<=', '≤') else 'gt' if op in ('大于', '超过', '>') else 'lt'
result['weightFilter'] = {'metric': 'scrapWeight' if '报废' in weight['kind'] else 'totalWeight', 'stage': stage, 'operator': op, 'value': float(weight['value'])}
return result
def select_rows(result, selection):
selection = validate_selection(selection)
rows = result.get('rows')
count = result.get('rowCount')
if (result.get('kind') != 'quality_metrics' or result.get('view') != selection['dimension']
or not isinstance(rows, list) or type(count) is not int or count != len(rows)
or result.get('rowsTruncated') is not False):
raise ValueError('汇总分组读取不完整,不能据此筛选排名')
def number(value):
return type(value) in (float, int) and math.isfinite(value)
weight_filter = selection.get('weightFilter')
eligible, missing, keys = [], 0, set()
for row in rows:
key = row.get('key')
if not isinstance(key, str) or not key or key in keys:
raise ValueError('分组键缺失或重复,不能稳定排序')
keys.add(key)
stage = selection['rateStage']
rate = (row.get(stage) or {}).get('rate') if stage else row.get('compositeRate')
# No-production fallback rates are not measured yields.
coverage = row.get('coverage') or {}
measured = coverage.get(stage + 'Rows', 0) > 0 if stage and stage + 'Rows' in coverage else True
if not number(rate) or rate < 0 or not measured:
missing += 1
continue
if weight_filter:
stages = [weight_filter['stage']] if weight_filter['stage'] in ('black', 'bright') else ['black', 'bright']
values = [(row.get(s) or {}).get(weight_filter['metric']) for s in stages]
compare = {'gt': operator.gt, 'gte': operator.ge, 'lt': operator.lt, 'lte': operator.le}[weight_filter['operator']]
matches = [number(v) and compare(v, weight_filter['value']) for v in values]
if not (any(matches) if weight_filter['stage'] == 'any' else all(matches)):
continue
eligible.append((rate, key, row))
eligible.sort(key=lambda item: (item[0] if selection['direction'] == 'asc' else -item[0], item[1]))
selected = [dict(copy.deepcopy(row), rank=index + 1, selectedRate=rate) for index, (rate, key, row) in enumerate(eligible[:selection['limit']])]
return {'selection': selection, 'rows': selected, 'sourceRows': count, 'eligibleRows': len(eligible),
'missingRateRows': missing, 'returnedRows': len(selected), 'completeSource': True,
'basis': '先按正式口径汇总全部授权分组,再按汇总重量条件筛选,按正式合格率排序,最后取Top N;不重算率,不平均产品率。'}
def project_selection(result, task):
selection = (task or {}).get('aggregateSelection')
if not selection or result.get('kind') != 'quality_metrics' or result.get('view') != selection['dimension']:
return result
selected = select_rows(result, selection)
# The signed original keeps every field. Give the model only the requested
# ranking columns, rather than 100 copies of defects and period summaries.
rows = []
for row in selected['rows']:
compact = {key: row[key] for key in ('key', 'label', 'rank', 'selectedRate') if key in row}
for stage in ('black', 'bright'):
compact[stage] = {key: (row.get(stage) or {}).get(key) for key in ('rate', 'totalWeight', 'scrapWeight')}
rows.append(compact)
view = {key: value for key, value in result.items() if key not in ('summary', 'trend', 'trendCount', 'trendTruncated')}
return {**view, 'rows': rows, 'rowsReturned': selected['returnedRows'],
'selectionDisplay': {key: value for key, value in selected.items() if key != 'rows'}}
def ranking_chart(document, task, turn_id):
from analysis_visuals import _chart, _chart_scope
selection = task['aggregateSelection']
result = document['result']; selected = select_rows(result, selection); rows = selected['rows']
if not rows:
return None
stage = selection['rateStage']; label = {'black': '黑片', 'bright': '亮片', None: '综合'}[stage]
# Split long rankings into readable native SVG panels, retaining every row.
charts = []
for offset in range(0, len(rows), 25):
page = rows[offset:offset + 25]
charts.append(_chart(turn_id=turn_id, chart_type='bar', title=f'{label}合格率排名 · {offset+1}—{offset+len(page)}',
subtitle=f"{'由低到高' if selection['direction'] == 'asc' else '由高到低'} · 符合条件 {selected['eligibleRows']} 项 · 展示 {len(rows)} 项",
unit='ratio', categories=[{'key': row['key'], 'label': f"{row['rank']}. {row.get('label') or row['key']}"} for row in page],
series=[{'key': stage+'_rate' if stage else 'composite_rate', 'label': label + '合格率', 'values': [row['selectedRate'] for row in page]}],
evidence_refs=[document['evidenceRef']], required_fields=([stage+'_good_weight', stage+'_total_weight'] if stage else ['black_good_weight', 'black_total_weight', 'bright_good_weight', 'bright_total_weight']),
scope=_chart_scope(result.get('filters', {}), result['releaseId'], result['period']), period=result['period'],
metric=stage+'_rate' if stage else 'composite_rate', stage=stage, dimension=selection['dimension'], section_key=selection['dimension'],
coverage={'returnedRows': len(page), 'totalRows': selected['eligibleRows'], 'truncated': selected['returnedRows'] < selected['eligibleRows']},
note=selected['basis']))
return charts