fix(check): 修留出集暴露的 5 处（规则集 2026.11-r3）
- 响应方式：「逐条响应」要说的是要求 / 参数 / 偏离，声明函填写说明不算；「否则视为无效响应」不算默认接受 - 偏离栏按「是不是明确满足」判断：偏离栏、响应情况栏一起看，其他短写法请人核对（最高重要）；表体序号认「技-86」等 - 另报价、附加条件按结构判断（另一个不同的总价金额、保留 / 另行 / 调整报价等） - 空白栏：「精确到年月日」不算没填 - 比对时折叠异体字（戶→户等），原文照存 - 记录决策 E2：先修再跑国动真实投标文件 开发集召回 75%、规矩投标文件问题数不变；留出集参考数字见 check-eval §12（已看过，仅参考）。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
YANG JIANKUAN
@@ -1,9 +1,11 @@
|
||||
"""偏离栏直读（m3-plan §2.1「偏离表直取」，check-eval §8）：模型判「已响应」的条目，看它引的证据句所在的表格行，
|
||||
同一行有单元格写着「偏离 / 负偏离 / 部分满足 / 未完全响应」这类字样的，改判负偏离。
|
||||
"""偏离栏直读（m3-plan §2.1「偏离表直取」，check-eval §8、§11）：模型判「已响应」的条目，看它引的证据句所在的表格行，
|
||||
偏离栏（按表头认出那一栏）写的不是明确的「无偏离 / 满足」，就改判负偏离。
|
||||
|
||||
开发集上负偏离未说明（D06）漏掉的，多数是响应格照抄了要求（「满足，……」）、偏离格写着「偏离」，
|
||||
模型只看了响应格——提示词已经要求读整行，模型还是会漏，这里用确定的规则补上。
|
||||
单元格要整格是这些字样才算（「如不满足……」这类正文不算）；改判后证据带上那一格，出处可回溯。
|
||||
- 偏离栏的值按「是不是明确满足」判断，不靠枚举不满足的说法：留出集写「不完全一致 / 与要求有出入 / 接近要求」，
|
||||
开发集的枚举一条都没认出（D06 92% → 25%）。明确写了不满足的（负偏离、部分满足……）照级别报；其余短写法
|
||||
（不超过 12 个字）请人核对，级别最高到「重要」；长段文字（偏离说明）不在这里判；
|
||||
- 认不出表头时退回整行找「负偏离 / 部分满足」这类明确写法；
|
||||
- 改判后证据带上那一格，出处可回溯。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
@@ -19,6 +21,85 @@ NEGATIVE = re.compile(
|
||||
r"^(未完全响应|不完全响应|部分响应|未响应|不响应|负偏离|有偏离|偏离|不满足|部分满足|未满足|不符合|部分符合)"
|
||||
r"([（(].*)?[。；]?$"
|
||||
)
|
||||
CLEAR_OK = re.compile(
|
||||
r"^(无|无偏离|没有偏离|不偏离|正偏离|优于要求|优于|完全响应|完全满足|满足|响应|已响应|符合|完全符合|一致|完全一致"
|
||||
r"|无差异|—+|-+|/|／)([（(].*)?[。；]?$"
|
||||
)
|
||||
DEV_HEAD = re.compile(r"偏离|响应情况|是否响应|响应结论|是否满足|满足情况|符合情况")
|
||||
YES_NO_HEAD = re.compile(r"是否(响应|满足|符合)")
|
||||
DEVIATE_HEAD = re.compile(r"是否偏离")
|
||||
SHORT = 12
|
||||
# 表体每行开头的序号：「12」「技-86」「C-3」「（一）」（风格 B / C 和不少真实文件用自己的编号）
|
||||
SERIAL = re.compile(r"(?:[A-Za-z一-鿿]{1,2}[-－])?\d{1,4}[.、]?|[（(]?[一二三四五六七八九十]{1,3}[)）、.]?")
|
||||
|
||||
|
||||
def classify(text: str, head: str = "") -> str:
|
||||
"""偏离 / 响应栏的一格：ok 明确满足、negative 明确不满足、unclear 其他短写法（请人核对）、other 长段说明或空。
|
||||
head 是这一栏的表头：「是否偏离」栏写「是」是偏离，「是否满足」栏写「否」是不满足。"""
|
||||
t = norm(text)
|
||||
if not t:
|
||||
return "other"
|
||||
if DEVIATE_HEAD.search(head) and t in ("是", "是。"):
|
||||
return "negative"
|
||||
if YES_NO_HEAD.search(head) and t in ("否", "否。", "×"):
|
||||
return "negative"
|
||||
if (YES_NO_HEAD.search(head) and t in ("是", "√")) or (DEVIATE_HEAD.search(head) and t in ("否", "否。")):
|
||||
return "ok"
|
||||
if NEGATIVE.match(t):
|
||||
return "negative"
|
||||
if CLEAR_OK.match(t):
|
||||
return "ok"
|
||||
return "unclear" if len(t) <= SHORT else "other"
|
||||
|
||||
|
||||
def is_dev_head(text: str) -> bool:
|
||||
t = norm(text)
|
||||
return (
|
||||
len(t) <= 30 and bool(DEV_HEAD.search(t)) and "说明" not in t
|
||||
) # 「投标文件中的响应情况（完全响应/未完全响应）」
|
||||
|
||||
|
||||
def judge_row(columns: list[tuple[list[Sentence], str]]) -> tuple[str, list[Sentence]]:
|
||||
"""一行里几栏（偏离栏、响应情况栏）合起来的结论：任一栏明确不满足就是 negative（「投标响应情况」写「部分响应」、
|
||||
偏离栏却写「符合」，开发集 S08-B）；否则偏离栏是 unclear 就是 unclear；返回结论和依据的那几段。"""
|
||||
kinds = [(judge(cells, head), cells) for cells, head in columns if cells]
|
||||
for want in ("negative", "unclear"):
|
||||
hit = next((cells for k, cells in kinds if k == want), None)
|
||||
if hit is not None:
|
||||
return want, hit
|
||||
return "ok" if kinds else "other", []
|
||||
|
||||
|
||||
NEG_WORDS = re.compile(r"负偏离|部分(响应|满足|符合)|未完全|不完全|不满足|未满足|不符合|未响应|不响应|有偏离")
|
||||
NEG_START = re.compile(r"^(" + NEG_WORDS.pattern + ")")
|
||||
# 「基本一致 / 大致符合 / 接近满足」不是明确满足，要请人核对
|
||||
OK_WORDS = re.compile(r"(?<!基本)(?<!大致)(?<!大体)(?<!接近)(?<!部分)(无偏|正偏|满足|完全响应|符合|一致)")
|
||||
|
||||
|
||||
def judge(cells: list[Sentence], head: str) -> str:
|
||||
"""一行里这一栏的几段合起来判断。窄栏会把「无偏离」折成「无偏」「离」，无框线的表还可能把相邻行的
|
||||
同一栏拼进来，所以按字样判断而不是整格匹配：有明确不满足的字样就是 negative；否则有「无偏 / 满足 /
|
||||
符合」就是 ok；一个字的碎片不判。只有「偏离」栏才把其他短写法算作 unclear；「响应情况」这类栏写的
|
||||
常是响应内容（「支持」「3年」），只认明确的不满足。"""
|
||||
pieces = [t for c in cells if (t := norm(c.text))]
|
||||
joined = "".join(pieces)
|
||||
if not joined:
|
||||
return "other"
|
||||
if joined in ("偏离", "偏离。") and not all(c.cell for c in cells):
|
||||
return "other" # 跨页、无框线的行抽字时「无偏离」会被拆成「无」和「偏离」（negative_cell 同一道限制）
|
||||
# 按段判断：短的一段里有不满足的字样，或长的一段开头就是不满足（「满足，……不符合……时预警」是正常响应）
|
||||
if any((len(t) <= SHORT and NEG_WORDS.search(t)) or NEG_START.match(t) for t in pieces):
|
||||
return "negative"
|
||||
by_head = classify(joined, head)
|
||||
if by_head in ("negative", "ok") and (DEVIATE_HEAD.search(head) or YES_NO_HEAD.search(head)):
|
||||
return by_head
|
||||
if OK_WORDS.search(joined):
|
||||
return "ok"
|
||||
if len(joined) <= 1:
|
||||
return "other"
|
||||
if by_head == "unclear" and "偏离" not in head:
|
||||
return "other"
|
||||
return by_head
|
||||
|
||||
|
||||
def negative(text: str) -> bool:
|
||||
@@ -32,30 +113,79 @@ def negative_cell(s: Sentence) -> bool:
|
||||
return negative(t) and (t not in ("偏离", "偏离。") or s.cell is not None)
|
||||
|
||||
|
||||
class Grid:
|
||||
"""投标文件按页分好的行：找一格所在的行，往上（同一页，再看上一页）找表头里的偏离栏。"""
|
||||
|
||||
def __init__(self, docs: list[BidDoc]) -> None:
|
||||
self.pages: dict[tuple[UUID, int], list[list[Sentence]]] = {}
|
||||
self.where: dict[int, tuple[UUID, int, int]] = {}
|
||||
for d in docs:
|
||||
for no, page in layout.pages(d.sentences).items():
|
||||
rows = layout.rows(page)
|
||||
self.pages[(d.id, no)] = rows
|
||||
for i, row in enumerate(rows):
|
||||
for c in row:
|
||||
self.where[id(c)] = (d.id, no, i)
|
||||
|
||||
def row(self, s: Sentence) -> list[Sentence]:
|
||||
at = self.where.get(id(s))
|
||||
return self.pages[(at[0], at[1])][at[2]] if at else []
|
||||
|
||||
def dev_columns(self, s: Sentence) -> list[tuple[list[Sentence], str]] | None:
|
||||
"""这一行的偏离栏、响应情况栏：[(这一栏的几段, 表头文字)]；找不到表头返回 None。"""
|
||||
at = self.where.get(id(s))
|
||||
if at is None:
|
||||
return None
|
||||
doc_id, no, i = at
|
||||
above = self.pages[(doc_id, no)][:i][::-1] + self.pages.get((doc_id, no - 1), [])[::-1]
|
||||
for head in above:
|
||||
if head and SERIAL.fullmatch(norm(head[0].text)):
|
||||
continue # 表体行（要求栏里也可能写着「……的响应情况」）
|
||||
heads = [(n, c) for n, c in enumerate(head) if c.cell and is_dev_head(c.text)]
|
||||
if not heads:
|
||||
continue
|
||||
cols = {f"c{n}": layout.span(c) for n, c in enumerate(head)}
|
||||
row = self.pages[(doc_id, no)][i]
|
||||
return [
|
||||
([c for c in row if c is not s and layout.column(c, cols) == f"c{n}"], norm(h.text)) for n, h in heads
|
||||
]
|
||||
return None
|
||||
|
||||
|
||||
def direct_read(results: dict[UUID, Result], docs: list[BidDoc]) -> int:
|
||||
"""就地改判，返回改判的条数。"""
|
||||
rows: dict[int, list[Sentence]] = {}
|
||||
for d in docs:
|
||||
for page in layout.pages(d.sentences).values():
|
||||
for row in layout.rows(page):
|
||||
for c in row:
|
||||
rows[id(c)] = row
|
||||
grid = Grid(docs)
|
||||
changed = 0
|
||||
for r in results.values():
|
||||
if r.verdict not in ("responded", "no_value") or not r.evidence:
|
||||
continue
|
||||
hit: tuple[BidDoc, Sentence] | None = None
|
||||
hit: tuple[BidDoc, Sentence, str, str] | None = None
|
||||
for d, s in r.evidence:
|
||||
cell = next((c for c in rows.get(id(s), []) if c is not s and negative_cell(c)), None)
|
||||
if not s.cell:
|
||||
continue # 正文里的证据句：没有偏离栏
|
||||
columns = grid.dev_columns(s)
|
||||
if columns is not None:
|
||||
kind, cells = judge_row(columns)
|
||||
if kind in ("negative", "unclear"):
|
||||
hit = (d, cells[0], kind, norm("".join(c.text for c in cells)))
|
||||
break
|
||||
continue
|
||||
cell = next((c for c in grid.row(s) if c is not s and negative_cell(c)), None) # 认不出表头
|
||||
if cell is not None:
|
||||
hit = (d, cell)
|
||||
hit = (d, cell, "negative", norm(cell.text))
|
||||
break
|
||||
if hit is None:
|
||||
continue
|
||||
d, cell, kind, text = hit
|
||||
before, r.verdict = r.verdict, "negative_deviation"
|
||||
if not any(x is hit[1] for _, x in r.evidence):
|
||||
r.evidence = [*r.evidence, hit][-3:]
|
||||
r.detail = {**r.detail, "direct_read": norm(hit[1].text), "model_verdict": before}
|
||||
if not any(x is cell for _, x in r.evidence):
|
||||
r.evidence = [*r.evidence, (d, cell)][-3:]
|
||||
r.detail = {
|
||||
**r.detail,
|
||||
"direct_read": text,
|
||||
"model_verdict": before,
|
||||
**({"deviation_unclear": True} if kind == "unclear" else {}),
|
||||
}
|
||||
changed += 1
|
||||
return changed
|
||||
|
||||
|
||||
@@ -77,6 +77,12 @@ def spec(result: Result, item: dict[str, Any]) -> Spec | None:
|
||||
if result.verdict == "not_found":
|
||||
text = f"在投标文件（{scope}）中没有找到对这条要求的响应。" if scope else "没有找到对这条要求的响应。"
|
||||
return Spec("unresponded", sev, f"未找到响应：{name}", text, base)
|
||||
if result.verdict == "negative_deviation" and result.detail.get("deviation_unclear"):
|
||||
text = (
|
||||
f"这一行的偏离栏写的是「{result.detail.get('direct_read')}」，不是明确的「无偏离 / 满足」，"
|
||||
"请核对是否满足这条要求。"
|
||||
)
|
||||
return Spec("negative_deviation", "important" if sev == "fatal" else sev, f"偏离栏需核对：{name}", text, base)
|
||||
if result.verdict == "negative_deviation":
|
||||
cell = result.detail.get("direct_read")
|
||||
text = (
|
||||
|
||||
@@ -12,7 +12,7 @@ import re
|
||||
from collections import Counter
|
||||
from dataclasses import dataclass
|
||||
|
||||
from bid_doctor.document.pipeline.text_layer import Sentence
|
||||
from bid_doctor.document.pipeline.text_layer import Sentence, fold
|
||||
|
||||
NUM_UNIT = re.compile(r"(\d+(?:\.\d+)?)\s*([A-Za-z]+|[个台套项条次人天年月日分秒毫兆核位路口])?")
|
||||
LATIN = re.compile(r"[A-Za-z][A-Za-z0-9\-]*")
|
||||
@@ -21,6 +21,7 @@ K1, B = 1.5, 0.75
|
||||
|
||||
|
||||
def tokens(text: str) -> list[str]:
|
||||
text = fold(text)
|
||||
out = [(n + (u or "")).lower() for n, u in NUM_UNIT.findall(text)]
|
||||
out += [w.lower() for w in LATIN.findall(text)]
|
||||
for run in CJK.findall(text):
|
||||
|
||||
@@ -553,9 +553,14 @@ def _lower_near(
|
||||
|
||||
|
||||
ALT = re.compile(
|
||||
r"(备选|替代|选择性|可选)(方案)?(报价|价格|总价)|优惠后?的?(报价|价格|总价|价)|折扣后?的?(报价|价格|总价|价)|"
|
||||
r"方案[一二三四五1-5][^，。；]{0,6}(报价|总价|价格)|第[二三][种套个份]?(报价|方案)|(让利|下浮|降价)后?的?(报价|价格|总价)"
|
||||
r"(备选|替代|选择性|可选)[^，。；]{0,6}(报价|价格|总价)|(优惠|折扣|让利|下浮|降价|最终)[^，。；]{0,4}(报价|价格|总价|价)|"
|
||||
r"方案[一二三四五1-5][^，。；]{0,6}(报价|总价|价格)|第[二三][种套个份]?(报价|方案)"
|
||||
)
|
||||
# 附条件的价格：「如……，可将总报价调整为 × 元」
|
||||
CONDITIONAL = re.compile(r"(如|若|如果|倘若|假如)[^。；]{0,40}(报价|总价|价格)[^。；]{0,8}(调整|降|下调|变更|改|调)为")
|
||||
PRICE_WORD = re.compile(r"总报价|投标报价|投标总价|响应报价|磋商报价|报价|总价")
|
||||
NOT_PRICE = re.compile(r"合同金额|合同价|中标金额|成交金额|业绩|注册资本|营业收入|资产总额|预算|限价|保证金|单价|分项")
|
||||
OTHER_SHARE = Decimal("0.3") # 另一个价格至少是投标总价的三成，才可能是「另一个报价」
|
||||
ALT_FATAL = re.compile(r"备选|替代|选择性|可选|方案[二三四五2-5]|第[二三]")
|
||||
NEG = re.compile(r"不|无|未|拒绝")
|
||||
RESP_PREFIX = re.compile(r"^(完全响应|已响应|响应|满足|无偏离|正偏离|同意|接受)[，,：:。、\s]*")
|
||||
@@ -582,20 +587,27 @@ class Tender:
|
||||
def check_quotes(
|
||||
docs: list[BidDoc], tender: Tender, tables: list[PriceTable], totals: list[Total], hits: list[Hit]
|
||||
) -> None:
|
||||
"""另外给的报价：「备选方案报价」「方案二总价」后面跟着金额（致命）；「优惠价 / 让利后」跟着的金额和投标总价
|
||||
不一样的（重要：多半是另报了一个价，也可能只是说明）。金额等于投标总价（开标一览表或分项合计）的、否定句、
|
||||
招标文件原话都不算。"""
|
||||
"""另外给的报价，按结构判断（check-eval §11：按措辞枚举在留出集上一条都没认出）：一句话里有报价字样或「备选 /
|
||||
优惠 / 让利」「如……调整为」这类说法，又有一个和投标总价（开标一览表或分项合计）不同、且不低于总价三成的金额
|
||||
（金额在前在后都算）。「备选 / 替代 / 方案二 / 第二套」按致命报，其余按重要报（也可能只是说明，请人核对）。
|
||||
否定句、招标文件原话、业绩合同金额、限价、保证金这类不算。"""
|
||||
values = {t.value for t in totals} | {t.total[1] for t in tables if t.total}
|
||||
ref = max(values) if values else None
|
||||
for d in docs:
|
||||
for s in d.sentences:
|
||||
m = ALT.search(s.text)
|
||||
if not m:
|
||||
amounts = money(s.text)
|
||||
if not amounts or tender.restated(s.text):
|
||||
continue
|
||||
amounts = money(s.text[m.end() :])
|
||||
clause = re.split(r"[，。；,;]", s.text[: m.start()])[-1]
|
||||
if not amounts or NEG.search(clause) or tender.restated(s.text) or set(amounts) & values:
|
||||
m = ALT.search(s.text) or CONDITIONAL.search(s.text)
|
||||
priced = PRICE_WORD.search(s.text) and not NOT_PRICE.search(s.text)
|
||||
if not (m or priced):
|
||||
continue
|
||||
fatal = bool(ALT_FATAL.search(m.group(0)))
|
||||
if m and NEG.search(re.split(r"[，。；,;]", s.text[: m.start()])[-1]):
|
||||
continue
|
||||
other = [a for a in amounts if a not in values and (a >= ref * OTHER_SHARE if ref else a >= 10_000)]
|
||||
if not other:
|
||||
continue
|
||||
fatal = bool(m and ALT_FATAL.search(m.group(0)))
|
||||
hits.append(
|
||||
Hit(
|
||||
"PRICE.MULTIPLE_QUOTES",
|
||||
@@ -746,7 +758,14 @@ def _filled(text: str) -> list[tuple[str, str]]:
|
||||
|
||||
BLANKS = (
|
||||
(re.compile(r"[_＿]{3,}"), "下划线处没有填"),
|
||||
(re.compile(r"(?<![\d〇零一二三四五六七八九十])(?<!出生)(?<!成立)(?<!注册)年\s*月\s*日"), "日期没有填"),
|
||||
# 「精确到年月日」「格式为年月日」「显示年月日时分」是说明，不是没填（留出集，check-eval §11）
|
||||
(
|
||||
re.compile(
|
||||
r"(?<![\d〇零一二三四五六七八九十])(?<!出生)(?<!成立)(?<!注册)(?<!精确到)(?<!格式为)(?<!显示)(?<!格式)"
|
||||
r"年\s*月\s*日(?!\s*[时\d])"
|
||||
),
|
||||
"日期没有填",
|
||||
),
|
||||
(re.compile(r"[（(]\s*[）)](?=\s*[（(][^）)]{2,12}[）)])"), "括号里没有填"),
|
||||
(re.compile(r"[:：]\s*[（(]\s*[）)]"), "括号里没有填"),
|
||||
(
|
||||
@@ -945,15 +964,22 @@ CONDITIONS = (
|
||||
),
|
||||
re.compile(r"按(实际|实际发生|实际需求|实际使用)[^。；，]{0,10}(结算|计算|计价|收费|计费)"),
|
||||
re.compile(r"(费用|费)[^。；，]{0,6}由(采购人|甲方|贵方|用户|业主)[^。；，]{0,4}(承担|负责|支付|另付)"),
|
||||
re.compile(r"另行(计算|计取|收取|收费|结算|报价|协商|商定|商议|约定|确定)|另议|另计"),
|
||||
re.compile(r"另行(计算|计取|收取|收费|结算|报价|协商|商定|商议|约定|确定|签订|商洽)|另议|另计"),
|
||||
re.compile(r"以(实际|最终|现场|实际发生)[^。；，]{0,12}为准"),
|
||||
re.compile(r"以[^。；，]{0,16}(另行|实际|最终|现场|届时|我方)[^。；，]{0,16}为准"),
|
||||
re.compile(r"保留[^。；，]{0,12}(权利|调整|变更)"),
|
||||
re.compile(r"(调整|变更|上调|上浮)[^。；，]{0,4}(报价|价格|单价|总价)"),
|
||||
re.compile(r"(价格|市场价|原材料|元器件)[^。；，]{0,10}(波动|上涨|变化)"),
|
||||
re.compile(r"(价格|报价|单价|总价)[^。；]{0,10}(随|根据|按照?)[^。；]{0,12}(调整|浮动|变动)"),
|
||||
re.compile(r"仅供参考"),
|
||||
re.compile(r"(须|需|需要|要求)[^。；]{0,6}(预付|预付款|首付款)"),
|
||||
)
|
||||
COMMERCIAL = re.compile(r"报价|价格|价款|总价|单价|费用|税|付款|支付|结算|数量|工程量|到货|交货|工期|质保|合同")
|
||||
COMMERCIAL = re.compile(
|
||||
r"报价|价格|价款|总价|单价|费用|税|付款|支付|结算|数量|工程量|到货|交货|交付|供货|工期|服务期|质保|售后|协议|合同"
|
||||
)
|
||||
NOT_CONDITION = re.compile(
|
||||
r"(不|无|未|免|不再|无需|不得|决不|不会)[^。；，]{0,4}(另行|另计|另议)|无效|否则|不予|不接受|免费|不影响|不增加"
|
||||
r"|(不|不得|不会|决不)以任何理由"
|
||||
)
|
||||
|
||||
|
||||
@@ -984,28 +1010,13 @@ def check_conditions(docs: list[BidDoc], tender: Tender, hits: list[Hit]) -> Non
|
||||
)
|
||||
|
||||
|
||||
RESP_HEAD = re.compile(r"(响应情况|是否响应|响应结论|偏离情况|是否偏离|偏离/响应|响应/偏离|是否满足|符合或偏离)")
|
||||
NEG_RESP = re.compile(
|
||||
r"^(未完全响应|不完全响应|部分响应|未响应|不响应|负偏离|有偏离|不满足|部分满足|未满足|不符合|部分符合)"
|
||||
r"([（(].*)?[。；]?$"
|
||||
)
|
||||
YES_NO_HEAD = re.compile(r"是否(响应|满足|符合)")
|
||||
DEVIATE_HEAD = re.compile(r"是否偏离")
|
||||
FATAL_TABLE = re.compile(r"(符合性|实质性|资格)")
|
||||
SERIAL = re.compile(r"\d{1,3}")
|
||||
|
||||
|
||||
def _negative(cell: str, head: str) -> bool:
|
||||
t = norm(cell)
|
||||
if NEG_RESP.match(t) or deviation.negative(t):
|
||||
return True
|
||||
if YES_NO_HEAD.search(head) and t in ("否", "×"):
|
||||
return True
|
||||
return bool(DEVIATE_HEAD.search(head)) and t == "是"
|
||||
SERIAL = deviation.SERIAL
|
||||
|
||||
|
||||
def check_response_tables(docs: list[BidDoc], hits: list[Hit]) -> None:
|
||||
"""响应表（表头有「响应情况 / 偏离情况」一栏）：这一栏写了「未完全响应」「负偏离」「否」之类，或者空着。
|
||||
"""响应表（表头有「响应情况 / 偏离」一栏）：这一栏写了「未完全响应」「负偏离」「否」之类，或者空着；
|
||||
写的不是明确的「无偏离 / 满足」的其他短写法（「不完全一致」「接近要求」）请人核对，按重要报（deviation.classify）。
|
||||
一条记录从序号行开始，到下一个序号行为止（垂直居中的格子、折行可能落在没有序号的行里）。
|
||||
表名、表头写「符合性 / 实质性 / 资格」的写了不响应按致命报，其余按重要报；「空着」可能是版面没认准，按重要报。
|
||||
能对上清单项的，响应匹配也会报，写入前按投标侧句子去重（drop_covered）。"""
|
||||
@@ -1013,18 +1024,18 @@ def check_response_tables(docs: list[BidDoc], hits: list[Hit]) -> None:
|
||||
for _, page in sorted(pages_of(d).items()):
|
||||
rows = rows_of(page)
|
||||
for i, row in enumerate(rows):
|
||||
at = next((n for n, c in enumerate(row) if RESP_HEAD.search(norm(c.text)) and len(c.text) <= 40), None)
|
||||
if at is None or len(row) < 3:
|
||||
if row and SERIAL.fullmatch(norm(row[0].text)):
|
||||
continue # 以序号开头的是表体（要求栏里也可能写着「……的响应情况」）
|
||||
heads = [(f"c{n}", norm(c.text)) for n, c in enumerate(row) if deviation.is_dev_head(c.text)]
|
||||
if not heads or len(row) < 3:
|
||||
continue
|
||||
head = norm(row[at].text)
|
||||
cols = {f"c{n}": _span(c) for n, c in enumerate(row)}
|
||||
want = f"c{at}"
|
||||
context = norm("".join(c.text for r in rows[max(0, i - 2) : i + 1] for c in r))
|
||||
fatal = bool(FATAL_TABLE.search(context))
|
||||
records: list[list[Sentence]] = []
|
||||
for r in rows[i + 1 :]:
|
||||
numbered = len(r) >= 2 and SERIAL.fullmatch(norm(r[0].text)) is not None
|
||||
if not numbered and RESP_HEAD.search(norm("".join(c.text for c in r))):
|
||||
if not numbered and any(deviation.is_dev_head(c.text) for c in r):
|
||||
break # 下一张表的表头（表体每行都以序号开头，要求栏里也可能写着「……的响应情况」）
|
||||
if numbered:
|
||||
records.append(list(r))
|
||||
@@ -1033,22 +1044,38 @@ def check_response_tables(docs: list[BidDoc], hits: list[Hit]) -> None:
|
||||
elif records:
|
||||
records[-1] += r
|
||||
for rec in records:
|
||||
resp = [c for c in rec if _col(c, cols) == want]
|
||||
columns = [([c for c in rec if _col(c, cols) == want], head) for want, head in heads]
|
||||
resp = [c for cells, _ in columns for c in cells]
|
||||
others = [c for c in rec[1:] if not any(c is x for x in resp)]
|
||||
if not others:
|
||||
continue
|
||||
req = others[0].text
|
||||
bad = next((c for c in resp if _negative(c.text, head)), None)
|
||||
if resp and bad is None:
|
||||
kind, basis = deviation.judge_row(columns)
|
||||
joined = norm("".join(c.text for c in basis))
|
||||
bad = basis[0] if kind == "negative" else None
|
||||
unclear = basis[0] if kind == "unclear" else None
|
||||
if resp and bad is None and unclear is None:
|
||||
continue
|
||||
if bad is not None:
|
||||
if bad is None and unclear is not None:
|
||||
hits.append(
|
||||
Hit(
|
||||
"FORM.RESPONSE_NEGATIVE",
|
||||
"negative_deviation",
|
||||
"important",
|
||||
f"响应表里写了「{joined}」",
|
||||
f"「{req[:40]}」这一行写的是「{joined}」，不是明确的「无偏离 / 满足」，请核对。",
|
||||
[(d, c) for c in rec],
|
||||
f"{d.no}:{target(req)}",
|
||||
)
|
||||
)
|
||||
elif bad is not None:
|
||||
hits.append(
|
||||
Hit(
|
||||
"FORM.RESPONSE_NEGATIVE",
|
||||
"negative_deviation",
|
||||
"fatal" if fatal else "important",
|
||||
f"响应表里写了「{norm(bad.text)}」",
|
||||
f"「{req[:40]}」这一行写的是「{norm(bad.text)}」。"
|
||||
f"响应表里写了「{joined}」",
|
||||
f"「{req[:40]}」这一行写的是「{joined}」。"
|
||||
+ (
|
||||
"表名写着符合性 / 实质性审查，这类要求不满足通常会导致投标无效，请核对。"
|
||||
if fatal
|
||||
|
||||
@@ -37,7 +37,7 @@ from bid_doctor.project import service as projects
|
||||
from bid_doctor.project.models import Project
|
||||
|
||||
CHECK_TASK = "check_project"
|
||||
RULE_SET_VERSION = "2026.11-r2" # 规则集有版本，检查批次记下当时用的版本（架构 §3.7）；r2：加分项 / 单价上限
|
||||
RULE_SET_VERSION = "2026.11-r3" # 规则集有版本，检查批次记下当时用的版本（架构 §3.7）；r3：留出集暴露的 5 处
|
||||
WAIVE_WINDOW = timedelta(hours=24)
|
||||
ERROR_MESSAGES = {
|
||||
"bid_text_missing": "投标文件里几乎读不出文字（可能都是扫描件），扫描件识别还没有上线",
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
"""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」；
|
||||
- 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」（要说的是要求、参数、偏离）；
|
||||
- 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。
|
||||
@@ -26,8 +26,13 @@ from bid_doctor.document.models import SourceDocument
|
||||
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)")
|
||||
# 「逐条响应」要说的是要求、参数、偏离：声明函填写说明「逐一、准确填写产品名称、厂名……」（招标文件的固定套话）不算
|
||||
# （留出集 S03 因此把「未列入偏离的视为接受」判成冲突，规矩的投标文件报了一千多个提示，check-eval §11）
|
||||
EXPLICIT_SUBJECT = re.compile(r"要求|条款|参数|指标|偏离|规格|需求|配置|技术|商务")
|
||||
# 「视为」后面是无效、不响应的，说的是不响应的后果，不是默认接受（「否则视为无效响应」，留出集 S09）
|
||||
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"(否则|未列入|未列出|未列明|除[^。；]{0,30}外)[^。；]{0,25}(视为|视作|认为)(?![^。；]{0,3}(无效|不|未|非))"
|
||||
r"[^。；]{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)")
|
||||
@@ -58,7 +63,8 @@ def _scope(i: int, sentences: list[Sentence]) -> set[str]:
|
||||
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))
|
||||
exp = bool(EXPLICIT.search(s.text) and EXPLICIT_SUBJECT.search(s.text))
|
||||
imp = bool(IMPLICIT.search(s.text))
|
||||
if not exp and not imp:
|
||||
continue
|
||||
for c in _scope(i, sentences):
|
||||
|
||||
@@ -39,6 +39,19 @@ RADICAL_SUPPLEMENT = dict(
|
||||
)
|
||||
|
||||
|
||||
# 同一个字的异体码位：有的字体把「户」输出成「戶」（U+6236），招标侧、投标侧字形不一致时检索和引文核对会对不上
|
||||
# （留出集渲染出的 PDF 就是这样，check-eval §11）。只在比对时折叠（norm），原文照存，和已存的文字层保持一致
|
||||
VARIANTS = str.maketrans(dict(zip("戶戸內靑淸眞卽旣敎", "户户内青清真即既教", strict=True)))
|
||||
|
||||
|
||||
def fold(s: str) -> str:
|
||||
"""比对用：异体码位、CJK 兼容汉字（U+F900–FAFF）换成常用写法。"""
|
||||
s = s.translate(VARIANTS)
|
||||
if any("\uf900" <= c <= "\ufaff" for c in s):
|
||||
s = "".join(unicodedata.normalize("NFKC", c) if "\uf900" <= c <= "\ufaff" else c for c in s)
|
||||
return s
|
||||
|
||||
|
||||
def _char(t: str) -> str:
|
||||
"""有的 PDF（如 Chromium 生成的）把汉字映射成康熙部首或部首补充区的码位（「目」成了 U+2F6C）。
|
||||
只把这两个区换回常用汉字；全角标点等其他字符保持原样，否则引文和原文对不上。"""
|
||||
@@ -369,4 +382,5 @@ def continuations(sentences: list[Sentence]) -> dict[str, Sentence]:
|
||||
|
||||
|
||||
def norm(s: str) -> str:
|
||||
return WS.sub("", s)
|
||||
"""比对用的文本：去空白、折叠异体字（fold）。"""
|
||||
return fold(WS.sub("", s))
|
||||
|
||||
Reference in New Issue
Block a user