Pair black and bright defect evidence for priority products

This commit is contained in:
2026-10-24 10:53:33 +08:00
parent 55327df5e4
commit edef747e3b
6 changed files with 252 additions and 7 deletions
+27 -7
View File
@@ -259,18 +259,38 @@ def synthesize(*, base_url: str, api_key: str, model: str, question: str,
request = Request(base_url.rstrip('/') + '/chat/completions',
data=json.dumps(body, ensure_ascii=False).encode(),
headers={'Content-Type': 'application/json', 'Authorization': 'Bearer ' + api_key}, method='POST')
def verified_fallback():
# Focus selection is presentation-only. If its model response is malformed,
# render the already verified facts with a deterministic default selection.
# Never promote missing server evidence to a complete answer.
selected = [candidates[0]['id'] for candidates in pools.values() if candidates]
missing = list(dict.fromkeys(report.get('missingEvidence', [])))[:10]
return {'answer':render_report(report,selected,missing),
'evidenceSufficient':not missing,'missingEvidence':missing,
'modelDeclaredEvidenceSufficient':None,'completionAdjusted':True,
'model':model,'usage':{},'finishReason':'selection_fallback',
'selectedFactIds':selected,'factCount':len(report['facts']),
'selectionAdjusted':True,'selectionFallback':True}
try:
with (opener or _open_model)(request, timeout=FINAL_TIMEOUT_SECONDS) as response:
raw = response.read(2 * 1024 * 1024 + 1)
if cancel_event.is_set(): raise FinalizationUnavailable('已取消')
if len(raw) > 2 * 1024 * 1024: raise ValueError('oversize')
document = json.loads(raw)
try:
document = json.loads(raw)
except (ValueError, TypeError):
return verified_fallback()
if not isinstance(document,dict):return verified_fallback()
if document.get('model') not in (None, model): raise ValueError('model changed')
choice = document['choices'][0]
answer = choice['message']['content']
if choice.get('finish_reason') != 'stop' or not isinstance(answer, str) or not answer.strip():
raise ValueError('incomplete final answer')
parsed=json.loads(answer)
try:
choice = document['choices'][0]
answer = choice['message']['content']
if choice.get('finish_reason') != 'stop' or not isinstance(answer, str) or not answer.strip():
return verified_fallback()
parsed=json.loads(answer)
except (ValueError, TypeError, KeyError, IndexError, AttributeError):
return verified_fallback()
if not isinstance(parsed,dict):return verified_fallback()
selected=[]
selection_adjusted=False
for field,candidates in pools.items():
@@ -283,7 +303,7 @@ def synthesize(*, base_url: str, api_key: str, model: str, question: str,
selected.extend(valid)
sufficient=parsed.get('evidenceSufficient');missing=parsed.get('missingEvidence')
if (not isinstance(sufficient,bool) or not isinstance(missing,list)
or any(not isinstance(x,str) for x in missing)):raise ValueError('invalid evidence completion')
or any(not isinstance(x,str) for x in missing)):return verified_fallback()
missing=list(dict.fromkeys(item.strip() for item in missing if item.strip()))[:10]
# 模型的孤立 false 不能证明数据缺失;具体缺项仍阻止完成,调用方另行检查服务端覆盖。
completion_adjusted = not sufficient and not missing
+67
View File
@@ -0,0 +1,67 @@
"""Plan missing black/bright source-defect queries for the same product scope.
Only current-turn, release-bound evidence is passed in by the runtime. This
module never estimates a missing stage or reuses another product's evidence.
"""
from __future__ import annotations
import json
import re
from typing import Any
def _single_stage_requested(question: str, context: Any) -> bool:
if isinstance(context, dict) and context.get("stage") in ("black", "bright"):
return True
text = question if isinstance(question, str) else ""
if re.search(r"(?:只|仅)(?:看|查|分析|要|关注)?(?:黑片|亮片)", text):
return True
stages = [stage for label, stage in (("黑片", "black"), ("亮片", "bright")) if label in text]
return len(stages) == 1 and not re.search(r"整体|总体|综合|全部|两阶段|重点产品|重点规格", text)
def plan_missing_product_defect_stages(
documents: list[dict[str, Any]], question: str, context: Any = None,
) -> list[dict[str, Any]]:
"""Return the missing counterpart queries; identical scope/period is required."""
if _single_stage_requested(question, context):
return []
groups: dict[str, dict[str, Any]] = {}
for document in documents:
result = document.get("result") if isinstance(document, dict) else None
if not isinstance(result, dict) or result.get("kind") != "quality_fine_defects":
continue
if result.get("dimension") != "defect" or result.get("stage") not in ("black", "bright"):
continue
filters = result.get("filters")
product = filters.get("product") if isinstance(filters, dict) else None
if not isinstance(product, str) or not product.strip():
continue
if result.get("category") is not None or result.get("defect") is not None:
continue
release = result.get("releaseId")
period = result.get("period")
comparison = result.get("comparisonPeriod")
if (not isinstance(release, str) or not release
or not isinstance(period, dict) or (comparison is not None and not isinstance(comparison, dict))):
continue
sort = result.get("sort")
if sort not in ("scrap_delta", "current_scrap"):
continue
identity = json.dumps([release, period, comparison, filters, sort], ensure_ascii=False, sort_keys=True)
group = groups.setdefault(identity, {"stages": set(), "source": result})
group["stages"].add(result["stage"])
queries = []
for group in groups.values():
if group["stages"] == {"black", "bright"}:
continue
source = group["source"]
other = "bright" if "black" in group["stages"] else "black"
args = {"mode": "defects", "releaseId": source["releaseId"],
"period": source["period"], "filters": source["filters"],
"dimension": "defect", "stage": other, "sort": source["sort"], "limit": 20}
if source.get("comparisonPeriod") is not None:
args["comparisonPeriod"] = source["comparisonPeriod"]
queries.append(args)
return queries
+43
View File
@@ -1980,6 +1980,49 @@ def run_turn(payload: dict[str, Any], *, stdout: Any = sys.stdout, stdin: Any =
actor_root=actor_root, cancel_event=cancel_event,
timeout=480 if model_name == "mimo-v2.6-flash" else TURN_TIMEOUT_SECONDS,
)
# The model may investigate one stage of a priority product and forget
# the other. Complete the same release/period/filter pair before the
# evidence-based answer and charts are assembled. Never borrow a result
# from another turn or infer a missing stage from the first one.
pairable_stop = status in ("COMPLETED", "PARTIAL") or (
status == "FAILED" and stop_reason in ("length", "timeout", "超时停止")
and stop_succeeded is True)
if pairable_stop and not cancel_event.is_set():
try:
from analysis_visual_evidence import load_visual_evidence
from nao_product_defect_pairing import plan_missing_product_defect_stages
from nao_quality_tools import QualityTools
documents = load_visual_evidence(actor_root / "evidence", request["releaseId"],
exclude_files=prior_evidence_files, budget_seconds=2.0)
missing_stages = plan_missing_product_defect_stages(
documents, request["question"], request["context"])
if missing_stages:
pairing_api = DelegatedApi(access_path, runtime_root=manager.users_root,
expected_actor_id=request["actorId"],
expected_session_id=request["sessionId"], timeout=5).login()
quality_tools = QualityTools(api_factory=lambda: pairing_api,
evidence_dir=actor_root / "evidence")
for arguments in missing_stages:
if cancel_event.is_set():
break
stage_name = "黑片" if arguments["stage"] == "black" else "亮片"
emit({"type": "progress", "message": f"正在补查同规格的{stage_name}具体缺陷…"})
try:
paired = quality_tools.quality_data(arguments)
except (ValueError, OSError, RuntimeError):
emit({"type": "progress", "message": f"{stage_name}具体缺陷补查未成功,结论将保留该阶段的数据限制。"})
continue
call = {"tool": "quality_data", "toolCallId": "pair-" + paired["evidenceRef"],
"label": f"补查{stage_name}具体缺陷", "status": "completed",
"startedAt": datetime.now(timezone.utc).isoformat(),
"finishedAt": datetime.now(timezone.utc).isoformat(),
"outputSummary": f"同规格、同期间补查完成;依据 {paired['evidenceRef']}"}
accumulator.tool_calls.append(call)
accumulator.record_event("tool_result", tool=call["tool"],
toolCallId=call["toolCallId"], label=call["label"],
status="completed", text=call["outputSummary"])
except (ValueError, OSError, RuntimeError):
emit({"type": "progress", "message": "阶段配对检查未完成,已有查询结果仍按原范围呈现。"})
from nao_evidence_finalizer import eligible, evidence_summaries, synthesize, coverage_sufficient, FinalizationUnavailable
finalized = False
evidence_completion = None