feat(check): 响应方式识别（C12）和「这条要不要投标人回答」；评测集纠错
- checklist/response_mode.py：按类别（技术 / 商务）识别招标文件的响应方式（逐条响应 / 默认接受），规则优先； 两种写法冲突按逐条，认不出按逐条并写进覆盖范围。 - check/applicability.py：无效情形（★ / 斜体下划线标注的条款除外）、日期与递交、没要求附材料的评分标准、 无需提供证明或由采购人查询的资格，不逐条找响应，写进覆盖范围；要求附证明材料按原文判断，中标后才交的不算。 - 默认接受的要求没找到响应：一般要求记 implicit_accepted、不出问题；实质性要求记 implicit_review、提示级 人工核对；要附证明材料的不免（迁移 0006）。 - match.py：修同一行单元格又是别的条目候选句时的 KeyError（每条清单项单独记候选句，句子从总表里查）。 - 评测集：注入器同一要求的所有表格行一起改；散文、附件、格式文件里也响应了的要求不挂漏响应类问题； 错位响应必须是另一个参数；复核判不成立的单元不再挂；regen.py 按固定种子重新生成。 开发集 17 份：规矩投标文件报出的问题 639 → 229，误报 21.3 → 12.4 / 100，召回 68% → 65%，致命级 20 → 21 / 32 （docs/architecture/check-eval.md §4）。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
YANG JIANKUAN
56
server/src/bid_doctor/check/applicability.py
Normal file
56
server/src/bid_doctor/check/applicability.py
Normal file
@@ -0,0 +1,56 @@
|
||||
"""每条清单项要不要在投标文件里找响应、按什么响应方式判（决策 C12，`docs/architecture/check-eval.md` §2.2 第 ② 项）。
|
||||
|
||||
开发集上规矩的投标文件报出的问题，近一半落在不需要投标人回答的条目上（无效情形的列举、日期、评分标准、
|
||||
采购人自己查询的资格）。这些条目不逐条找响应，写进检查覆盖范围：
|
||||
- 无效情形、评审规定（只有「无效」类别、不是标注出来的条款）：由规则检查（第 2 步）和人工核对覆盖；
|
||||
- 日期与递交：由规则检查（落款日期、有效期）和需人工核对覆盖；
|
||||
- 评分标准本身：没写要附什么材料的不找；写了的按证明材料找；
|
||||
- 资格条件里写明「无需提供」或由采购人查询（信用中国）的不找。
|
||||
其余按类别取招标文件的响应方式：技术参数用技术类，其余用商务类。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from dataclasses import dataclass
|
||||
from typing import Any
|
||||
|
||||
from bid_doctor.checklist.response_mode import Mode
|
||||
|
||||
# 要求投标时附证明材料（读标的 evidence_required 常常是空的，按原文判断）；中标后 / 签合同时才交的不算
|
||||
EVIDENCE = re.compile(r"(提供|附|出具)[^。；]{0,20}(证明|报告|截图|复印件|扫描件|证书|合同|承诺|声明|社保|发票)")
|
||||
AWARD = re.compile(r"(中标后|成交后|签订合同(前|时)|合同签订(前|时))")
|
||||
NO_PROOF = re.compile(r"无需提供|不需提供|无须提供|采购人[^。；]{0,12}(查询|核查|查验)|信用中国|信用记录")
|
||||
SKIP_LABELS = {
|
||||
"condition": "无效情形和评审规定",
|
||||
"schedule": "日期与递交要求",
|
||||
"scoring": "评分标准（没要求附证明材料）",
|
||||
"no_proof": "资格条件（写明无需提供证明，或由采购人查询）",
|
||||
}
|
||||
|
||||
|
||||
@dataclass(frozen=True, slots=True)
|
||||
class Decision:
|
||||
match: bool
|
||||
mode: str = "explicit" # explicit / implicit
|
||||
skip: str | None = None # 不找响应的原因（SKIP_LABELS 的键）
|
||||
mode_status: str = "found"
|
||||
evidence: bool = False # 要求投标时附证明材料：默认接受也不能免（没附就是没附）
|
||||
|
||||
|
||||
def decide(item: dict[str, Any], modes: dict[str, Mode]) -> Decision:
|
||||
cats = set(item.get("categories") or [])
|
||||
text = f"{item.get('title') or ''}{item.get('requirement_text') or ''}"
|
||||
if cats == {"invalidity"} and item.get("importance_basis") != "marker":
|
||||
return Decision(False, skip="condition") # 「……的，投标无效」这类情形；★ / 斜体下划线标注的条款是要求，照常找
|
||||
if "schedule" in cats and not cats & {"tech_param", "qualification", "scoring"}:
|
||||
return Decision(False, skip="schedule")
|
||||
evidence = bool(item.get("evidence_required") or EVIDENCE.search(text)) and not AWARD.search(text)
|
||||
if cats <= {"scoring"} and not evidence:
|
||||
return Decision(False, skip="scoring")
|
||||
if "qualification" in cats and "tech_param" not in cats and NO_PROOF.search(text):
|
||||
return Decision(False, skip="no_proof")
|
||||
mode = modes.get("technical" if "tech_param" in cats else "commercial")
|
||||
if mode is None:
|
||||
return Decision(True, evidence=evidence)
|
||||
return Decision(True, mode=mode.mode, mode_status=mode.status, evidence=evidence)
|
||||
@@ -52,8 +52,14 @@ def spec(result: Result, item: dict[str, Any]) -> Spec | None:
|
||||
name = str(item.get("title") or "")
|
||||
base = {"importance_basis": item.get("importance_basis"), **result.detail}
|
||||
scope = "、".join(result.detail.get("search_scope") or [])
|
||||
if result.verdict == "responded":
|
||||
if result.verdict in ("responded", "implicit_accepted"):
|
||||
return None
|
||||
if result.verdict == "implicit_review":
|
||||
text = (
|
||||
"招标文件规定未列入偏离表的视为接受，投标文件里没有找到对这条要求的明确响应。"
|
||||
"这是实质性要求，请人工核对所投内容确实满足。"
|
||||
)
|
||||
return Spec("manual_review", "hint", f"请人工核对：{name}", text, {**base, "mode": "implicit"})
|
||||
if result.verdict == "not_found" and result.evidence_status == "unverified":
|
||||
claimed = result.detail.get("claimed_pages") or []
|
||||
where = f"{_pages(claimed)}有响应" if claimed else "有响应"
|
||||
|
||||
@@ -74,6 +74,18 @@ class Stats:
|
||||
self.verify_failures[code] = self.verify_failures.get(code, 0) + 1
|
||||
|
||||
|
||||
def apply_modes(results: dict[UUID, Result], items: dict[UUID, dict[str, Any]], modes: dict[UUID, Any]) -> None:
|
||||
"""按招标文件的响应方式改结论（决策 C12）：默认接受（implicit）的要求没找到响应不算问题，记「按招标文件默认接受」；
|
||||
实质性要求出提示级「请人工核对」，不出致命。找到了不满足的响应照常报。modes：清单项 → applicability.Decision。"""
|
||||
for item_id, d in modes.items():
|
||||
r = results.get(item_id)
|
||||
if r is None or d.mode != "implicit" or r.verdict != "not_found" or d.evidence:
|
||||
continue # 要附证明材料的要求，默认接受也免不了
|
||||
r.verdict = "implicit_review" if items[item_id].get("importance") == "substantive" else "implicit_accepted"
|
||||
r.evidence_status, r.reason_code = "none", None
|
||||
r.detail = {**r.detail, "mode": "implicit"}
|
||||
|
||||
|
||||
def short_sid(doc: BidDoc, s: Sentence) -> str:
|
||||
return f"d{doc.no}p{s.page_no}s{s.seq}"
|
||||
|
||||
|
||||
@@ -20,16 +20,18 @@ from sqlalchemy.ext.asyncio import AsyncSession
|
||||
from bid_doctor.ai.models import AiCall
|
||||
from bid_doctor.ai.providers.base import TransientError
|
||||
from bid_doctor.billing import service as billing
|
||||
from bid_doctor.check import findings, match
|
||||
from bid_doctor.check import applicability, findings, match
|
||||
from bid_doctor.check.models import CheckItemResult, CheckRun, CheckRunDocument, Finding, FindingCitation
|
||||
from bid_doctor.check.service import RULE_SET_VERSION, open_fatal
|
||||
from bid_doctor.checklist import edit as checklist_edit
|
||||
from bid_doctor.checklist import response_mode
|
||||
from bid_doctor.checklist import service as checklist
|
||||
from bid_doctor.core import events
|
||||
from bid_doctor.core.db import TenantContext, tenant_tx
|
||||
from bid_doctor.core.errors import AppError
|
||||
from bid_doctor.core.ids import uuid7
|
||||
from bid_doctor.document import citations
|
||||
from bid_doctor.document import service as documents
|
||||
from bid_doctor.document.citations import document_label
|
||||
from bid_doctor.document.models import SourceDocument
|
||||
from bid_doctor.document.pipeline import normalize, text_layer
|
||||
@@ -136,6 +138,23 @@ async def _stages(ctx: TenantContext, run_id: UUID) -> None:
|
||||
snapshot = await checklist_edit.version_items(s, r.checklist_version_id)
|
||||
lots = await projects.lot_labels(s, r.project_id)
|
||||
pending = (await checklist.pending_counts(s, [r.project_id])).get(r.project_id, 0)
|
||||
tenders = (
|
||||
await s.execute(
|
||||
select(SourceDocument).where(
|
||||
SourceDocument.project_id == r.project_id,
|
||||
SourceDocument.role.in_(("tender", "addendum", "clarification")),
|
||||
SourceDocument.text_layer_key.is_not(None),
|
||||
)
|
||||
)
|
||||
).scalars()
|
||||
tender_sents = [
|
||||
x
|
||||
for d in tenders
|
||||
for x in text_layer.body_sentences(
|
||||
[text_layer.Page.from_json(pg) for pg in await documents.text_layer(s, ctx, d)]
|
||||
)
|
||||
]
|
||||
modes = response_mode.detect(tender_sents) # 招标文件按类别的响应方式（C12）
|
||||
bidding = set(r.lot_ids)
|
||||
items, other_lot = {}, 0
|
||||
for item_id, payload in snapshot.items():
|
||||
@@ -144,6 +163,12 @@ async def _stages(ctx: TenantContext, run_id: UUID) -> None:
|
||||
items[item_id] = payload
|
||||
else:
|
||||
other_lot += 1 # 不在所投包里的条目不查，写进覆盖范围（C6）
|
||||
decisions = {i: applicability.decide(p, modes) for i, p in items.items()}
|
||||
skipped: dict[str, int] = {}
|
||||
for d in decisions.values():
|
||||
if not d.match and d.skip:
|
||||
skipped[d.skip] = skipped.get(d.skip, 0) + 1
|
||||
items = {i: p for i, p in items.items() if decisions[i].match} # 不需要投标人回答的条目不逐条找响应
|
||||
last = [PROGRESS["match"]]
|
||||
|
||||
async def progress(ratio: float) -> None:
|
||||
@@ -153,6 +178,7 @@ async def _stages(ctx: TenantContext, run_id: UUID) -> None:
|
||||
await _stage(ctx, run_id, "match", pct)
|
||||
|
||||
results, stats = await match.run(ctx, run_id, items, bid_docs, progress)
|
||||
match.apply_modes(results, items, decisions)
|
||||
|
||||
await _stage(ctx, run_id, "rules") # 第 2 步：报价形式、形式要件、需人工核对
|
||||
await _stage(ctx, run_id, "verify") # 前两道校验在匹配时逐条做完；这里只是进度节点
|
||||
@@ -174,7 +200,7 @@ async def _stages(ctx: TenantContext, run_id: UUID) -> None:
|
||||
)
|
||||
).one()
|
||||
r.counts = await _counts(s, run_id)
|
||||
r.coverage = _coverage(r, pages_info, items, results, pending, other_lot, stats, tokens)
|
||||
r.coverage = _coverage(r, pages_info, items, results, pending, other_lot, stats, tokens, skipped, modes)
|
||||
await _publish(ctx, r, "running", "summarize", PROGRESS["summarize"])
|
||||
|
||||
|
||||
@@ -352,6 +378,8 @@ def _coverage(
|
||||
other_lot: int,
|
||||
stats: match.Stats,
|
||||
tokens: Any,
|
||||
skipped: dict[str, int],
|
||||
modes: dict[str, response_mode.Mode],
|
||||
) -> dict[str, Any]:
|
||||
"""检查覆盖范围：报告首页、查标页顶部都读这里（FR-M3-09、m3-plan §2.5）。"""
|
||||
undetermined = sum(x.verdict == "undetermined" for x in results.values())
|
||||
@@ -367,6 +395,19 @@ def _coverage(
|
||||
not_checked.append(f"{undetermined} 条没有完成判断（模型漏答或暂时不可用），已列为「请人工核对」")
|
||||
if scanned:
|
||||
not_checked.append(f"{scanned} 页没有文字层（扫描件），这些页上的内容没有检查")
|
||||
for reason, n in skipped.items():
|
||||
not_checked.append(f"{applicability.SKIP_LABELS[reason]} {n} 条：不在投标文件里逐条找响应")
|
||||
labels = {"technical": "技术要求", "commercial": "商务要求"}
|
||||
for m in modes.values():
|
||||
if m.status == "unknown":
|
||||
not_checked.append(f"招标文件没有写明{labels[m.category]}是否要求逐条响应，按逐条响应处理")
|
||||
elif m.status == "conflict":
|
||||
not_checked.append(
|
||||
f"招标文件对{labels[m.category]}的响应方式写法不一致（既要求逐条响应，又写明未列出的视为接受），按逐条响应处理"
|
||||
)
|
||||
implicit = sum(x.verdict in ("implicit_accepted", "implicit_review") for x in results.values())
|
||||
if implicit:
|
||||
checked.append(f"其中 {implicit} 条按招标文件「未列入偏离的视为接受」处理，没有复述不算问题")
|
||||
not_checked.append("报价形式、形式要件、资质匹配：规则检查还没有上线")
|
||||
if not r.checklist_confirmed:
|
||||
not_checked.append("清单未确认，检查结果以当前清单为准")
|
||||
@@ -383,6 +424,16 @@ def _coverage(
|
||||
"evidence_unverified": unverified,
|
||||
},
|
||||
"checklist_confirmed": r.checklist_confirmed,
|
||||
"skipped": skipped,
|
||||
"response_modes": [
|
||||
{
|
||||
"category": m.category,
|
||||
"mode": m.mode,
|
||||
"status": m.status,
|
||||
"basis": [{"page": x.page_no, "text": x.text[:120]} for x in (m.implicit + m.explicit)[:4]],
|
||||
}
|
||||
for m in modes.values()
|
||||
],
|
||||
"legal_regime": {"value": r.legal_regime, "rules_verified": r.legal_regime in REGIMES_VERIFIED},
|
||||
"rule_set_version": RULE_SET_VERSION,
|
||||
"prompt_versions": r.prompt_versions,
|
||||
|
||||
63
server/src/bid_doctor/checklist/response_mode.py
Normal file
63
server/src/bid_doctor/checklist/response_mode.py
Normal file
@@ -0,0 +1,63 @@
|
||||
"""招标文件对投标人怎么响应的规定（决策 C12，`evals/check/GUIDELINES.md` §2）：按类别（技术 / 商务）识别响应方式。
|
||||
|
||||
- explicit 逐条响应：「逐条 / 逐项作出响应」「逐一逐条响应填写」「逐条标明满足与否」；
|
||||
- implicit 默认接受：「否则将认为投标人接受」「未列入上表的视作完全接受」「除所列偏离外均视作理解和响应」
|
||||
「其余无偏离内容无需赘述」；
|
||||
- 两种写法都有（S07）按逐条处理，记 conflict，问题说明里引两处原文；一种都没有按逐条处理，记 unknown。
|
||||
规则优先、纯函数；结果带原文句子，查标写进检查覆盖范围。
|
||||
TODO(C12)：存进清单、在清单页显示并可改（人工确认点）；模型兜底规则没认出的写法。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from dataclasses import dataclass, field
|
||||
|
||||
from bid_doctor.document.pipeline.text_layer import Sentence
|
||||
|
||||
EXPLICIT = re.compile(r"逐(条|项|一)[^。；]{0,15}(响应|应答|标明|填写|填列|作出|做出|说明)")
|
||||
IMPLICIT = re.compile(
|
||||
r"(否则|未列入|未列出|未列明|除[^。；]{0,30}外)[^。；]{0,25}(视为|视作|认为)[^。；]{0,15}(接受|响应|理解)"
|
||||
r"|无需赘述"
|
||||
)
|
||||
TABLE_HINT = re.compile(r"(技术|商务)[^。；]{0,12}(偏离|响应)")
|
||||
LOOKBACK = 12 # 往前找几句，看这句说明属于哪张表（技术 / 商务）
|
||||
CATEGORIES = ("technical", "commercial")
|
||||
|
||||
|
||||
@dataclass(slots=True)
|
||||
class Mode:
|
||||
category: str # technical / commercial
|
||||
mode: str # explicit / implicit
|
||||
status: str # found / conflict / unknown
|
||||
explicit: list[Sentence] = field(default_factory=list[Sentence])
|
||||
implicit: list[Sentence] = field(default_factory=list[Sentence])
|
||||
|
||||
|
||||
def _scope(i: int, sentences: list[Sentence]) -> set[str]:
|
||||
"""这句说明管哪一类：句子自己写了「技术 / 商务」就按它，否则看前面几句里最近的表名；都没有就两类都管。"""
|
||||
own = {c for c, k in (("technical", "技术"), ("commercial", "商务")) if k in sentences[i].text}
|
||||
if own:
|
||||
return own
|
||||
for s in reversed(sentences[max(0, i - LOOKBACK) : i]):
|
||||
if m := TABLE_HINT.search(s.text):
|
||||
return {"technical" if m.group(1) == "技术" else "commercial"}
|
||||
return set(CATEGORIES)
|
||||
|
||||
|
||||
def detect(sentences: list[Sentence]) -> dict[str, Mode]:
|
||||
hits: dict[str, Mode] = {c: Mode(c, "explicit", "unknown") for c in CATEGORIES}
|
||||
for i, s in enumerate(sentences):
|
||||
exp, imp = bool(EXPLICIT.search(s.text)), bool(IMPLICIT.search(s.text))
|
||||
if not exp and not imp:
|
||||
continue
|
||||
for c in _scope(i, sentences):
|
||||
(hits[c].implicit if imp else hits[c].explicit).append(s)
|
||||
for m in hits.values():
|
||||
if m.explicit and m.implicit:
|
||||
m.mode, m.status = "explicit", "conflict"
|
||||
elif m.implicit:
|
||||
m.mode, m.status = "implicit", "found"
|
||||
elif m.explicit:
|
||||
m.mode, m.status = "explicit", "found"
|
||||
return hits
|
||||
Reference in New Issue
Block a user