eric/bid-doctor
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

feat(evals): 查标评测集生成器、注入、评分和只检索基线

Browse Source 
- build.py：从「招标文件档案」和读标清单生成规矩投标文件稿；技术参数按 op / value / unit 算满足的值，
  分项报价按分项上限比例分摊，落款日期取截止日前 3 天，偏离表按格式表头决定列（含设备名称列），
  读标漏抽的技术参数也补上响应；routes 支持「+」组合和 prose 散文响应。
- mutate.py：按词库注入要求项问题（单点 / 连续块，带不带标志词）和规则问题，写响应清单；
  符合性条件响应表的问题按表级记一个。
- score.py：按单元去重、类型族比对、读标漏抽单列、端到端召回、无法对照单列。
- run_check.py：加只检索不调模型的基线；render.py：扫描件按最后一次出现定位。
- vocab-dev.json：开发集注入词库（留出集的另行封存，不进 git）。

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
YANG JIANKUAN
2026-11-04 12:18:48 +08:00
parent 01ddc86281
commit a07663cbdb
6 changed files with 1070 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

445
evals/check/build.py Normal file
View File

@@ -0,0 +1,445 @@
"""从「招标文件档案」和读标清单生成规矩投标文件的结构化稿 spec.json（evals/check/GUIDELINES.md §1、§5）。
cd server && ../.venv/bin/python ../evals/check/build.py <招标文件> <投标文件编号> --style A|B [--checklist fix3]
规矩靠构造保证：偏离表、响应表、报价表、证明材料由脚本生成；技术参数有 op / value / unit 的，按要求算一个满足的值。
LLM 只写散文（档案里的 prose）和 B / C 风格的改写（rewrite.json，可选）。
## 招标文件档案 evals/out/check/tenders/<ID>.json（照着招标原文整理，不进 git）
{
"tender": "S02",
"project": {"name": "…", "code": "…", "purchaser": "…", "agency": "…", "deadline": "2026-12-01",
"validity_days": 90, "ceiling_total": "1514100.00", "lots": [{"no": "1", "ceiling": "…"}]},
"modes": [{"category": "tech_param", "scope": "一般参数", "mode": "implicit", "when": null,
"basis": [{"p": 76, "t": "否则将认为投标人接受"}]}],
"forms": [ // 按格式章的顺序
{"key": "bid_letter", "title": "投标函", "kind": "text", "paragraphs": ["致：{agency}", "…{project_name}…"]},
{"key": "price_summary", "title": "开标一览表", "kind": "price_summary", "header": ["项目名称", "投标总价（元）", "…"]},
{"key": "price_items", "title": "分项价格表", "kind": "price_items", "header": ["序号", "名称", "…"],
"items": [{"name": "汇聚交换机", "qty": 3, "unit": "台"}]},
{"key": "tech_dev", "title": "偏离表", "kind": "response_table", "route": "tech_row",
"header": ["序号", "招标文件的要求", "投标文件中的内容", "偏离", "说明"], "only_deviations": true},
{"key": "compliance", "title": "采购需求符合性条件响应表", "kind": "fixed_rows", "header": ["…"],
"rows": [["1", "“(一)数量要求”的响应情况", "见招标文件第六章《采购需求》中相关要求", "完全响应", ""]]},
{"key": "auth", "title": "授权委托书", "kind": "text", "paragraphs": ["…"]},
{"key": "evidence", "title": "资格证明文件", "kind": "attachments"}
],
"attachments": [{"key": "a1", "name": "营业执照（副本）", "scan": true, "lines": ["…"]}],
"routes": {"i12": "tech_row", "i40": "commercial_row", "i55": "attach:a1", "i60": "form:compliance", "i70": "na"},
"prose": [{"title": "技术方案", "paragraphs": ["…"]}]
}
routes 的取值可以用「+」连起来（tech_row+attach:a3），「prose」表示靠技术方案散文响应（prose:<节标题> 指定哪一节）。
占位符：{bidder} {credit_code} {legal_rep} {agent} {project_name} {project_code} {purchaser} {agency} {date}
{validity_days} {total} {total_cn}。`routes` 的键是读标清单里正式条目的序号（items.json 的下标，前缀 i）。
"""
from __future__ import annotations
import argparse
import datetime
import json
import pathlib
import re
import sys
from decimal import ROUND_HALF_UP, Decimal
from typing import Any
EVALS = pathlib.Path(__file__).resolve().parents[1]
sys.path.insert(0, str(EVALS))
import gold # noqa: E402
OUT = EVALS / "out" / "check"
BIDDER = {
"bidder": "江苏启衡信息技术有限公司（虚构）",
"credit_code": "91320100MA1XXXXX0Q",
"legal_rep": "周明远",
"agent": "陈立新",
"date": "2026年11月20日",
}
DIGITS = "零壹贰叁肆伍陆柒捌玖"
UNITS = ["", "拾", "佰", "仟"]
SECTIONS = ["", "万", "亿"]
def rmb_upper(amount: Decimal) -> str:
"""金额大写（元、角、分），用于开标一览表。"""
q = amount.quantize(Decimal("0.01"), rounding=ROUND_HALF_UP)
yuan, cents = divmod(int(q * 100), 100)
if yuan == 0:
head = "零元"
else:
groups, out = [], ""
while yuan:
groups.append(yuan % 10000)
yuan //= 10000
for gi in range(len(groups) - 1, -1, -1):
g = groups[gi]
if g == 0:
if out and not out.endswith("零"):
out += "零"
continue
part, zero = "", False
for di in range(3, -1, -1):
d = (g // 10**di) % 10
if d == 0:
zero = bool(part)
continue
if zero:
part += "零"
zero = False
part += DIGITS[d] + UNITS[di]
if out and g < 1000 and not out.endswith("零"):
out += "零"
out += part + SECTIONS[gi]
head = out.rstrip("零") + "元"
jiao, fen = divmod(cents, 10)
if not cents:
return head + "整"
return head + (DIGITS[jiao] + "角" if jiao else "零") + (DIGITS[fen] + "分" if fen else "")
def fill(text: str, ctx: dict[str, str]) -> str:
return re.sub(r"\{(\w+)\}", lambda m: ctx.get(m.group(1), m.group(0)), text)
def comply(spec: dict[str, Any], positive: bool) -> tuple[str, str]:
"""按要求算一个满足的响应值。返回 (响应文字, 偏离)。positive 时给正偏离（T05）。"""
param, op, value, unit = (spec.get(k) or "" for k in ("param", "op", "value", "unit"))
try:
v = Decimal(str(value))
except Exception: # noqa: BLE001
v = None
if v is not None and op in (">=", "<="):
better = (v * 2 if op == ">=" else v / 2).normalize() if positive else v.normalize()
return f"{param}：{better:f}{unit}", "正偏离" if positive else "无偏离"
if v is not None and op == "=":
return f"{param}：{v.normalize():f}{unit}", "无偏离"
return f"支持{param}" if param else "满足", "无偏离"
def load_items(tender: str, run: str) -> dict[str, dict[str, Any]]:
rows = json.loads((EVALS / "out" / "runs" / run / tender / "items.json").read_text())
by_sid = {gold.sid(s): s for s in gold.sentences(tender)}
out = {}
for i, r in enumerate(rows):
if r.get("zone") != "formal":
continue
sents = [by_sid[k] for k in r.get("sentences") or [] if k in by_sid]
out[f"i{i}"] = {
**r,
"text": "".join(s.text for s in sents),
"req": {"p": sents[0].page_no, "t": gold.norm(sents[0].text)[:24]} if sents else None,
}
return out
def unit_of(tender: str) -> dict[str, str]:
"""清单条目 → 标注单元：读标清单条目的句子和标注单元的句子有重叠就对上（和读标评分同一口径）。"""
units, _ = gold.units(tender)
by_sent: dict[str, list[str]] = {}
for u in units:
for s in u.sids:
by_sent.setdefault(s, []).append(u.id)
return by_sent # 句子编号 → 单元编号
COLUMN_HINTS = [ # 表头文字 → 字段（按顺序匹配，先匹配到的为准）
("no", ("序号", "序")),
("deviation", ("偏离",)),
("note", ("说明", "备注")),
("answer", ("响应", "投标文件中的内容", "应答", "投标人", "实际")),
("requirement", ("要求", "条款", "参数", "指标")),
("device", ("名称", "设备", "货物", "服务")),
]
def columns_of(header: list[str]) -> list[str]:
"""按格式的表头决定每一列放什么（S04 的偏离表第 2 列是「项目名称」）。认不出的列留空。"""
out: list[str] = []
for h in header:
field = next((f for f, keys in COLUMN_HINTS if any(k in h for k in keys) and f not in out), "")
out.append(field)
if "answer" not in out and "" in out:
out[out.index("")] = "answer"
return out
def _row(k: int, r: dict[str, Any], columns: list[str]) -> dict[str, Any]:
value = {
"no": str(k),
"device": (r.get("tp") or {}).get("device") or "",
"requirement": r["requirement"],
"answer": r["answer"],
"deviation": r["deviation"],
"note": "",
}
return {
"cells": [value.get(c, "") for c in columns],
"anchor": columns.index("answer") if "answer" in columns else len(columns) - 1,
"responds": r["responds"],
"req": r["req"],
"iid": r["iid"],
"tp": r["tp"],
}
def build(tender: str, bid: str, style: str, run: str) -> None:
profile = json.loads((OUT / "tenders" / f"{tender}.json").read_text())
items = load_items(tender, run)
sent_units = unit_of(tender)
rewrite_path = OUT / bid / "rewrite.json"
rewrite = json.loads(rewrite_path.read_text()) if rewrite_path.exists() else {}
proj = profile["project"]
routes: dict[str, str] = profile.get("routes", {})
rows_by_route: dict[str, list[dict[str, Any]]] = {}
attach_responds: dict[str, list[str]] = {}
form_responds: dict[str, list[str]] = {}
for iid, it in items.items():
units = sorted({u for s in it.get("sentences") or [] for u in sent_units.get(s, [])})
# 一条要求可以走几个去处，用「+」连起来（如 tech_row+attach:a3：表里一行，另附证明材料）
parts = routes.get(iid, "na").split("+")
for part in parts:
if part.startswith("attach:"):
attach_responds.setdefault(part[7:], []).extend(units)
elif part.startswith("form:"):
form_responds.setdefault(part[5:], []).extend(units)
elif part.startswith("prose"):
form_responds.setdefault("prose:" + part[6:], []).extend(units)
route = next((x for x in parts if x in ("tech_row", "commercial_row")), None)
if route is None:
continue
raw = (it.get("raw") or {}).get("tech_param") or {}
positive = style == "A" and len(rows_by_route.get(route, [])) % 9 == 4 # 每 9 行一个正偏离（T05）
req_text = it["text"] or it["title"]
numeric = raw.get("op") in (">=", "<=", "=") and raw.get("value")
if numeric:
answer, dev = comply(raw, positive)
else: # 功能类要求：复述型（风格 A）照常写「满足，……」
answer, dev = "满足，" + re.sub(r"^[\s★▲*＊\d.、（）()]+", "", req_text)[:120], "无偏离"
rw = rewrite.get(iid) or {}
rows_by_route.setdefault(route, []).append(
{
"iid": iid,
"requirement": rw.get("requirement", req_text),
"answer": rw.get("answer", answer),
"deviation": dev,
"responds": units,
"req": it["req"],
"tp": raw,
}
)
covered = {u for v in rows_by_route.values() for r in v for u in r["responds"]}
covered |= {u for v in attach_responds.values() for u in v} | {u for v in form_responds.values() for u in v}
na_units = sorted(
{
u
for iid, it in items.items()
if routes.get(iid, "na") == "na"
for s in it.get("sentences") or []
for u in sent_units.get(s, [])
}
- covered
)
# 读标清单漏抽、标注集里有的技术参数：也按规矩响应，免得规矩的投标文件里「无意漏响应」
gold_units, _ = gold.units(tender)
by_sid = {gold.sid(x): x for x in gold.sentences(tender)}
for u in gold_units:
if u.cat != "tech_param" or u.id in covered or u.id in na_units or not u.sids:
continue
sents = sorted((by_sid[k] for k in u.sids if k in by_sid), key=lambda x: (x.page_no, x.seq))
if not sents:
continue
text = "".join(x.text for x in sents)
rows_by_route.setdefault("tech_row", []).append(
{
"iid": None,
"requirement": text,
"answer": "满足：" + text,
"deviation": "无偏离",
"responds": [u.id],
"req": {"p": sents[0].page_no, "t": gold.norm(sents[0].text)[:24]},
"tp": {},
}
)
if proj.get("deadline"): # 落款日期取投标截止日前 3 天（写死的日期可能晚于截止日，规矩的投标文件不能带 P08）
dl = datetime.date.fromisoformat(str(proj["deadline"])[:10]) - datetime.timedelta(days=3)
bidder_date = f"{dl.year}年{dl.month}月{dl.day}日"
else:
bidder_date = BIDDER["date"]
total = Decimal(str(proj.get("ceiling_total") or "1000000")) * Decimal("0.93")
total = total.quantize(Decimal("1"))
ctx = {
**BIDDER,
"date": bidder_date,
"project_name": proj.get("name", ""),
"project_code": proj.get("code", ""),
"purchaser": proj.get("purchaser", ""),
"agency": proj.get("agency", ""),
"validity_days": str(proj.get("validity_days") or 90),
"total": f"{total:,.2f}",
"total_cn": rmb_upper(total),
}
blocks: list[dict[str, Any]] = [{"type": "title", "text": f"{proj.get('name', '')}投标文件"}]
price_items = next((f.get("items") for f in profile["forms"] if f["kind"] == "price_items"), []) or []
weights = [Decimal(str(x.get("qty") or 1)) for x in price_items] or [Decimal(1)]
# 各分项的合价：有分项 / 单价上限的按上限的比例分摊（总价是限价的 93%，所以每项都在上限以内）；没有的按数量分摊
ceilings = [Decimal(str(x["ceiling"])) if x.get("ceiling") else None for x in price_items]
if price_items and all(c is not None for c in ceilings):
base = [
c * w if x.get("ceiling_is_unit") else c for c, w, x in zip(ceilings, weights, price_items, strict=False)
] # type: ignore[operator]
else:
base = list(weights)
lines = [(total * b / sum(base)).quantize(Decimal("0.01")) for b in base]
lines[-1] = total - sum(lines[:-1])
unit_prices = [(ln / w).quantize(Decimal("0.01")) for ln, w in zip(lines, weights, strict=False)]
for n, f in enumerate(profile["forms"], start=1):
head = {"type": "heading", "text": f"{n}．{f['title']}", "new_page": n > 1, "form": f["key"]}
kind = f["kind"]
if kind == "text":
blocks += [
head,
*(
{
"type": "para",
"text": fill(p, ctx),
"responds": form_responds.get(f["key"], []),
"form": f["key"],
}
for p in f["paragraphs"]
),
]
blocks.append({"type": "sign", "lines": [f"投标人：{ctx['bidder']}（电子签章）", f"日期：{ctx['date']}"]})
elif kind == "price_summary":
header = f.get("header") or ["项目名称", "投标总价（小写，元）", "投标总价（大写）"]
cells = [ctx["project_name"], ctx["total"], ctx["total_cn"]] + [""] * max(0, len(header) - 3)
blocks += [
head,
{
"type": "table",
"header": header,
"rows": [{"cells": cells[: len(header)], "anchor": 1}],
"price_summary": True,
},
]
elif kind == "price_items":
rows = []
for k, (x, up, ln) in enumerate(zip(price_items, unit_prices, lines, strict=False), start=1):
rows.append(
{
"cells": [
str(k),
x["name"],
f"{x.get('qty', 1)}",
x.get("unit", ""),
f"{up:,.2f}",
f"{ln:,.2f}",
],
"anchor": 1,
}
)
header = ["序号", "名称", "数量", "单位", "单价（元）", "总价（元）"]
blocks += [
head,
{"type": "table", "header": header, "rows": rows, "price_items": True},
{"type": "para", "text": f"合计：{ctx['total']}元（{ctx['total_cn']}）"},
]
elif kind == "response_table":
src = rows_by_route.get(f.get("route", "tech_row"), [])
header = f.get("header") or ["序号", "招标文件的要求", "投标文件中的内容", "偏离", "说明"]
columns = columns_of(header)
if f.get("only_deviations"):
devs = [r for r in src if r["deviation"] != "无偏离"]
rows = [_row(k, r, columns) for k, r in enumerate(devs, 1)]
empty = {"no": "1", "requirement": "—", "answer": "无偏离", "deviation": "无"}
rows = rows or [{"cells": [empty.get(c, "") for c in columns], "anchor": columns.index("answer")}]
pool = [
_row(0, r, columns) for r in src if r["deviation"] == "无偏离"
] # 没列出的（默认接受）：注入时可以挑来写成偏离
else:
rows = [_row(k, r, columns) for k, r in enumerate(src, 1)]
pool = []
blocks += [
head,
{
"type": "table",
"header": header,
"rows": rows,
"form": f["key"],
"route": f.get("route", "tech_row"),
"only_deviations": bool(f.get("only_deviations")),
"columns": columns,
"pool": pool,
},
]
elif kind == "fixed_rows":
rows = [
{"cells": [fill(c, ctx) for c in r], "responds": form_responds.get(f["key"], []), "anchor": len(r) - 2}
for r in f["rows"]
]
blocks += [head, {"type": "table", "header": f["header"], "rows": rows, "form": f["key"], "fixed": True}]
elif kind == "attachments":
blocks.append(head)
for a in profile.get("attachments", []):
b = {
"type": "scan" if a.get("scan") else "para",
"text": fill(a["name"], ctx),
"responds": sorted(set(attach_responds.get(a["key"], []))),
"attach": a["key"],
}
if a.get("scan"):
b["lines"] = [fill(x, ctx) for x in a.get("lines", [])]
else:
b["text"] = fill(a["name"] + "：" + "；".join(a.get("lines", [])), ctx)
blocks.append(b)
else:
raise ValueError(f"未知的表单类型：{kind}")
for sec in profile.get("prose", []):
blocks.append({"type": "heading", "text": sec["title"], "new_page": True})
units = sorted(set(form_responds.get("prose:" + sec["title"], []) + form_responds.get("prose:", [])))
blocks += [
{"type": "para", "text": fill(p, ctx), "responds": units if i == 0 else [], "form": "prose"}
for i, p in enumerate(sec["paragraphs"])
]
spec = {
"bid": bid,
"tender": tender,
"style": style,
"project_name": proj.get("name", ""),
"bidder": BIDDER,
"volumes": [{"id": "v1", "kind": "whole", "header": None, "blocks": blocks}],
"price": {
"total": f"{total:.2f}",
"items": [
{"name": x["name"], "qty": x.get("qty", 1), "unit_price": f"{p:.2f}"}
for x, p in zip(price_items, unit_prices, strict=False)
],
},
"routes_used": {r: len(v) for r, v in rows_by_route.items()},
"na_units": na_units,
"deadline": proj.get("deadline"),
"ceiling_total": proj.get("ceiling_total"),
"validity_days": proj.get("validity_days"),
}
d = OUT / bid
d.mkdir(parents=True, exist_ok=True)
(d / "spec.json").write_text(json.dumps(spec, ensure_ascii=False, indent=1))
print(
f"{bid}：{sum(len(v) for v in rows_by_route.values())} 行响应，{len(profile.get('attachments', []))} 份附件，总价 {ctx['total']}"
)
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("tender")
ap.add_argument("bid")
ap.add_argument("--style", default="A")
ap.add_argument("--checklist", default="fix3")
a = ap.parse_args()
build(a.tender, a.bid, a.style, a.checklist)

390
evals/check/mutate.py Normal file
View File

@@ -0,0 +1,390 @@
"""往规矩的投标文件稿里注入问题，写出马虎的投标文件稿和响应清单（evals/check/GUIDELINES.md §3、§6.2）。
cd server && ../.venv/bin/python ../evals/check/mutate.py <规矩的投标文件> <输出的投标文件> --split dev [--seed 1] [--density high|low]
cd server && ../.venv/bin/python ../evals/check/mutate.py <规矩的投标文件> --clean --split dev # 只给规矩的投标文件写响应清单
- 词库：开发集用 evals/check/vocab-dev.json；留出集用封存的 evals/out/check/_sealed/vocab-holdout.json；
- 注入是确定性的（同一个 seed 结果相同），每个问题记 defect、variant、granularity、worded；
- 问题只挂在标注单元上；挂上的单元先记 reviewed: false，复核（决策 E1）后改成 true，评分只算复核过的；
- 响应方式按单元在投标文件里的响应位置推出来：逐条响应表 → explicit，只列偏离的表 → implicit，附件 → evidence，格式文件 → presence。
"""
from __future__ import annotations
import argparse
import copy
import json
import pathlib
import random
import re
import sys
from decimal import Decimal
from typing import Any
EVALS = pathlib.Path(__file__).resolve().parents[1]
sys.path.insert(0, str(EVALS))
sys.path.insert(0, str(EVALS / "check"))
import gold # noqa: E402
from build import rmb_upper # noqa: E402
OUT = EVALS / "out" / "check"
VOCAB = {"dev": EVALS / "check" / "vocab-dev.json", "holdout": OUT / "_sealed" / "vocab-holdout.json"}
UNIT_DOWN = {"TB": ("GB", 1024), "GB": ("MB", 1024), "Gbps": ("Mbps", 1000), "s": ("ms", 1000), "秒": ("毫秒", 1000)}
class Mutator:
def __init__(self, spec: dict[str, Any], vocab: dict[str, Any], rng: random.Random) -> None:
self.spec, self.vocab, self.rng = spec, vocab, rng
self.unit_defects: dict[str, dict[str, Any]] = {}
self.rule_defects: list[dict[str, Any]] = []
self.blocks = spec["volumes"][0]["blocks"]
# ---------- 定位 ----------
def tables(self, *, explicit: bool | None = None, fixed: bool = False) -> list[dict[str, Any]]:
out = []
for b in self.blocks:
if b.get("type") != "table" or not b.get("form"):
continue
if fixed:
if b.get("fixed"):
out.append(b)
continue
if b.get("fixed") or "only_deviations" not in b:
continue
if explicit is None or b["only_deviations"] != explicit:
out.append(b)
return out
def free(self, units: list[str]) -> bool:
return bool(units) and not any(u in self.unit_defects for u in units)
def mark(
self, units: list[str], defect: str, variant: str, granularity: str = "single", worded: bool = True
) -> None:
for u in units:
self.unit_defects[u] = {"defect": defect, "variant": variant, "granularity": granularity, "worded": worded}
def phrase(self, key: str, field: str) -> str:
return str(self.rng.choice(self.vocab[key][field]))
# ---------- 要求项问题 ----------
def d01_block(self) -> None:
"""连续漏掉 5～10 行（逐条响应表）。"""
for t in self.rng.sample(self.tables(explicit=True), k=len(self.tables(explicit=True))):
rows = t["rows"]
if len(rows) < 12:
continue
n = self.rng.randint(5, 10)
start = self.rng.randint(0, len(rows) - n)
chunk = rows[start : start + n]
if not all(self.free(r.get("responds") or []) for r in chunk if r.get("responds")):
continue
for r in chunk:
self.mark(r.get("responds") or [], "D01", "删连续一段", "block", worded=False)
del rows[start : start + n]
return
def row_ops(self, n: int) -> None:
"""逐条响应表里的单点问题：漏行、数值不满足、写明不支持、负偏离未说明、只写满足、对象错位。"""
cands = [(t, r) for t in self.tables(explicit=True) for r in t["rows"] if self.free(r.get("responds") or [])]
self.rng.shuffle(cands)
ops = ["D01", "D03", "D03", "D05", "D06", "D08", "D09"]
done = 0
for t, r in cands:
if done >= n:
break
if r not in t["rows"] or not self.free(r.get("responds") or []):
continue
# 从轮到的类型开始，这一行做不了就换下一种（软件类项目几乎没有数值参数，D03 常常做不了）
for k in range(len(ops)):
if self._apply(ops[(done + k) % len(ops)], t, r, cands):
done += 1
break
@staticmethod
def col(t: dict[str, Any], field: str) -> int | None:
cols = t.get("columns") or ["no", "requirement", "answer", "deviation", "note"]
return cols.index(field) if field in cols else None
def _apply(self, op: str, t: dict[str, Any], r: dict[str, Any], cands: list[Any]) -> bool:
ans, dev, note, req = (self.col(t, f) for f in ("answer", "deviation", "note", "requirement"))
if ans is None:
return False
tp = r.get("tp") or {}
units = r.get("responds") or []
if op == "D01":
t["rows"].remove(r)
self.mark(units, "D01", "删一行", worded=False)
elif op == "D03":
return self._unmet(r, tp, ans, dev)
elif op == "D05":
param = tp.get("param") or (r["cells"][req][:12] if req is not None else "该功能")
r["cells"][ans] = self.phrase("D05", "phrases").format(param=param)
self.mark(units, "D05", r["cells"][ans])
elif op == "D06":
if dev is None:
return False
r["cells"][dev] = self.phrase("D06", "deviations")
if note is not None:
r["cells"][note] = ""
self.mark(units, "D06", r["cells"][dev])
elif op == "D08":
if not tp.get("value"):
return False
r["cells"][ans] = self.phrase("D08", "phrases")
self.mark(units, "D08", r["cells"][ans], worded=False)
elif op == "D09":
others = [
x
for _, x in cands
if x is not r
and (x.get("tp") or {}).get("param")
and x["cells"][x.get("anchor", ans)] != r["cells"][ans]
]
if not others:
return False
o = self.rng.choice(others)
r["cells"][ans] = o["cells"][o.get("anchor", ans)]
self.mark(units, "D09", "另一条参数的响应", worded=False)
return True
def _unmet(self, r: dict[str, Any], tp: dict[str, Any], ans: int, dev: int | None) -> bool:
"""把响应值改成不满足；一部分换成更小的单位（换算后不满足）。不带标志词。"""
op, unit = tp.get("op"), tp.get("unit") or ""
try:
v = Decimal(str(tp.get("value")))
except Exception: # noqa: BLE001
return False
if op not in (">=", "<=") or v == 0:
return False
bad = (v / 2) if op == ">=" else (v * Decimal("1.6"))
variant = "低于下限" if op == ">=" else "高于上限"
if unit in UNIT_DOWN and self.rng.random() < 0.4:
small, k = UNIT_DOWN[unit]
text, variant = f"{(bad * k).normalize():f}{small}", "换算后不满足"
else:
text = f"{bad.normalize():f}{unit}"
r["cells"][ans] = f"{tp.get('param') or ''}：{text}"
if dev is not None and self.rng.random() < 0.5:
r["cells"][dev] = "无偏离" # 一半在偏离栏仍写「无偏离」（另一半保持原样，也是无偏离或正偏离）
self.mark(r.get("responds") or [], "D03", variant, worded=False)
return True
def implicit_deviation(self, n: int) -> None:
"""只列偏离的表：挑默认接受的要求写成偏离（D06 负偏离未说明 / D05 写明不支持）。"""
for t in self.tables(explicit=False):
pool = [r for r in t.get("pool") or [] if self.free(r.get("responds") or [])]
self.rng.shuffle(pool)
ans, dev, note, req = (self.col(t, f) for f in ("answer", "deviation", "note", "requirement"))
if ans is None or dev is None:
continue
for k, r in enumerate(pool[:n]):
row = copy.deepcopy(r)
if k % 2 == 0:
row["cells"][dev] = self.phrase("D06", "deviations")
if note is not None:
row["cells"][note] = ""
self.mark(row["responds"], "D06", row["cells"][dev])
else:
param = (row.get("tp") or {}).get("param") or (
row["cells"][req][:12] if req is not None else "该功能"
)
row["cells"][ans] = self.phrase("D05", "phrases").format(param=param)
row["cells"][dev] = "负偏离"
self.mark(row["responds"], "D05", row["cells"][ans])
if t["rows"] and "—" in t["rows"][0]["cells"]:
t["rows"].clear()
if self.col(t, "no") is not None:
row["cells"][self.col(t, "no")] = str(len(t["rows"]) + 1)
t["rows"].append(row)
def attachments(self, n: int, cats: dict[str, str]) -> None:
"""删掉附件：资格条件 → D11，评分项 → D13，其他 → D02。"""
cands = [b for b in self.blocks if b.get("attach") and self.free(b.get("responds") or [])]
self.rng.shuffle(cands)
for b in cands[:n]:
units = b.get("responds") or []
kinds = {cats.get(u) for u in units}
defect = "D11" if "qualification" in kinds else "D13" if "scoring" in kinds else "D02"
self.mark(units, defect, "附件缺失", worded=False)
self.blocks.remove(b)
def compliance(self) -> None:
for t in self.tables(fixed=True):
rows = [r for r in t["rows"] if self.free(r.get("responds") or [])]
if not rows:
continue
r = self.rng.choice(rows)
at = r.get("anchor", len(r["cells"]) - 2)
r["cells"][at] = self.phrase("D10", "cells")
# 符合性条件响应表的每一行没法对到单个标注单元（整张表共用一组），按表级问题记一个
self.rule("D10", r["cells"][at] or "空着", r["cells"][1] if len(r["cells"]) > 1 else "")
return
# ---------- 规则问题 ----------
def rule(self, kind: str, variant: str, snippet: str) -> None:
self.rule_defects.append(
{"id": f"{kind}-{len(self.rule_defects) + 1}", "type": kind, "variant": variant, "t": snippet}
)
def price_table(self) -> dict[str, Any] | None:
return next((b for b in self.blocks if b.get("price_summary")), None)
def p04(self) -> None:
t = self.price_table()
if not t or len(t["rows"][0]["cells"]) < 3: # 开标一览表没有大写金额栏（S04）：做不了
return
variant = self.phrase("P04", "variants")
total = Decimal(self.spec["price"]["total"])
wrong = {"drop_digit": total / 10, "extra_shi": total + 10, "decimal_shift": total / 100}.get(
variant, total + 1
)
t["rows"][0]["cells"][2] = rmb_upper(wrong.quantize(Decimal("0.01")))
self.rule("P04", variant, t["rows"][0]["cells"][2])
def p05(self) -> None:
t = next((b for b in self.blocks if b.get("price_items")), None)
if not t or not t["rows"]:
return
r = self.rng.choice(t["rows"])
r["cells"][-1] = f"{Decimal(r['cells'][-1].replace(',', '')) + 1000:,.2f}"
self.rule("P05", "分项总价多 1000 元", r["cells"][-1])
def p06(self) -> None:
fp = self.rng.choice(self.vocab["P06"]["foreign_projects"])
self.spec["volumes"][0]["header"] = fp["name"]
self.rule("P06", "页眉留旧项目名", fp["name"])
def text_paras(self, form_hint: str) -> list[dict[str, Any]]:
return [b for b in self.blocks if b.get("type") == "para" and form_hint in str(b.get("form") or "")]
def p07(self) -> None:
bidder = self.spec["bidder"]["bidder"]
paras = [b for b in self.blocks if b.get("type") == "para" and b.get("form") and bidder in b["text"]]
if paras:
b = self.rng.choice(paras)
blank = self.phrase("P07", "blanks")
b["text"] = b["text"].replace(bidder, blank, 1)
self.rule("P07", "投标人名称空", b["text"][:24])
def p08(self) -> None:
signs = [b for b in self.blocks if b.get("type") == "sign"]
if signs:
date = self.phrase("P08", "dates")
b = self.rng.choice(signs)
b["lines"] = [re.sub(r"日期：.*", f"日期：{date}", x) for x in b["lines"]]
self.rule("P08", date, date)
def p09(self) -> None:
days = str(self.spec.get("validity_days") or 90)
for b in self.blocks:
if b.get("type") == "para" and f"{days}天" in b["text"].replace(" ", ""):
new = str(self.phrase("P09", "days"))
b["text"] = b["text"].replace(f"{days}天", f"{new}天").replace(f"{days} 天", f"{new}天")
self.rule("P09", f"{days} → {new} 天", f"{new}天")
return
def p10(self) -> None:
letter = next((i for i, b in enumerate(self.blocks) if b.get("type") == "para" and b.get("form")), None)
if letter is not None:
text = self.phrase("P10", "phrases")
self.blocks.insert(letter + 1, {"type": "para", "text": text, "form": self.blocks[letter]["form"]})
self.rule("P10", text, text)
def p12(self) -> None:
t = self.price_table()
if t:
alt = f"{Decimal(self.spec['price']['total']) * Decimal('0.97'):,.2f}"
text = self.phrase("P12", "phrases").format(alt=alt)
self.blocks.insert(self.blocks.index(t) + 1, {"type": "para", "text": text})
self.rule("P12", text, text)
def unit_modes(spec: dict[str, Any]) -> dict[str, str]:
modes: dict[str, str] = {}
for b in spec["volumes"][0]["blocks"]:
if b.get("type") == "table" and "only_deviations" in b:
mode = "implicit" if b["only_deviations"] else "explicit"
for r in [*b["rows"], *(b.get("pool") or [])]:
for u in r.get("responds") or []:
modes.setdefault(u, mode)
elif b.get("attach"):
for u in b.get("responds") or []:
modes.setdefault(u, "evidence")
elif b.get("form"):
for r in b.get("rows") or [b]:
for u in r.get("responds") or []:
modes.setdefault(u, "presence")
return modes
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("clean")
ap.add_argument("out", nargs="?")
ap.add_argument("--clean", dest="manifest_only", action="store_true", help="只给规矩的投标文件写响应清单")
ap.add_argument("--split", required=True, choices=["dev", "holdout"])
ap.add_argument("--seed", type=int, default=1)
ap.add_argument("--density", default="high", choices=["high", "low"])
a = ap.parse_args()
spec = json.loads((OUT / a.clean / "spec.json").read_text())
tender = spec["tender"]
units, _ = gold.units(tender)
cats = {u.id: u.cat for u in units}
optional = {u.id for u in units if u.optional}
modes = unit_modes(spec) # 注入前算：被删掉的行也要知道它本来该怎么响应
m = Mutator(spec, json.loads(VOCAB[a.split].read_text()), random.Random(a.seed)) # noqa: S311 评测数据，不用于安全
if not a.manifest_only:
if a.density == "high":
m.d01_block()
m.row_ops(18)
m.implicit_deviation(4)
m.attachments(4, cats)
m.compliance()
for op in (m.p04, m.p05, m.p06, m.p07, m.p08, m.p09, m.p10, m.p12):
op()
else:
m.row_ops(4)
m.attachments(1, cats)
m.p04()
m.p07()
bid = a.clean if a.manifest_only else a.out
assert bid
manifest_units: dict[str, Any] = {}
na = set(spec.get("na_units") or [])
for u in units:
mode = modes.get(u.id) or ("na" if u.id in na else "unknown")
entry: dict[str, Any] = {"cat": u.cat, "mode": mode, "optional": u.id in optional}
if u.id in m.unit_defects:
entry.update({"expect": "defect", "reviewed": False, **m.unit_defects[u.id]})
elif mode == "implicit":
entry.update({"expect": "ok", "trap": "T06"})
else:
entry["expect"] = {"na": "na", "unknown": "unknown"}.get(mode, "ok")
manifest_units[u.id] = entry
manifest = {
"bid": bid,
"tender": tender,
"style": spec["style"],
"split": a.split,
"kind": "clean" if a.manifest_only else "sloppy",
"seed": a.seed,
"density": None if a.manifest_only else a.density,
"units": manifest_units,
"rule_defects": m.rule_defects,
}
d = OUT / bid
d.mkdir(parents=True, exist_ok=True)
spec["bid"] = bid
(d / "spec.json").write_text(json.dumps(spec, ensure_ascii=False, indent=1))
(d / "manifest.json").write_text(json.dumps(manifest, ensure_ascii=False, indent=1))
n_def = sum(1 for e in manifest_units.values() if e.get("expect") == "defect")
by: dict[str, int] = {}
for e in manifest_units.values():
if e.get("expect") == "defect":
by[e["defect"]] = by.get(e["defect"], 0) + 1
print(f"{bid}：要求项问题 {n_def} 个 {by}，规则问题 {len(m.rule_defects)} 个")
if __name__ == "__main__":
main()

11
evals/check/render.py
View File

@@ -132,12 +132,11 @@ def block_text(b: dict[str, Any], row: dict[str, Any] | None = None) -> str:
return str(b.get("text") or " ".join(b.get("lines") or []))
def locate(texts: dict[int, str], snippet: str) -> int | None:
def locate(texts: dict[int, str], snippet: str, last: bool = False) -> int | None:
"""last：取最后一次出现的页（附件在文件末尾，名称常和前面格式文件的标题相同）。"""
q = text_layer.norm(snippet)[:SNIPPET]
for p, t in texts.items():
if q and q in t:
return p
return None
hits = [p for p, t in texts.items() if q and q in t]
return (hits[-1] if last else hits[0]) if hits else None
def render(bid: str) -> None:
@@ -151,7 +150,7 @@ def render(bid: str) -> None:
page_html, header = volume_html(spec, vol)
pdf = render_pdf(page_html, header)
texts = page_texts(pdf)
scans = {p for b in vol["blocks"] if b.get("type") == "scan" and (p := locate(texts, b["text"]))}
scans = {p for b in vol["blocks"] if b.get("type") == "scan" and (p := locate(texts, b["text"], last=True))}
located: list[tuple[list[str], str]] = []
for b in vol["blocks"]:
if b.get("type") == "table":

27
evals/check/run_check.py
View File

@@ -14,6 +14,7 @@ import asyncio
import hashlib
import json
import pathlib
import re
import sys
from typing import Any
from uuid import UUID, uuid4
@@ -66,6 +67,29 @@ async def cached_run_json(
return data["data"]
ITEM = re.compile(r"^\[(I\d+)\] ")
CANDS = re.compile(r"^候选句：(.*)$")
FRAG = re.compile(r"^\[(d\d+p\d+s\d+)\]（[^）]*）(.*)$")
async def baseline_run_json(
ctx: Any, *, job_type: str, job_id: Any, task: str, prompt_id: str, version: int, user: str
) -> dict[str, Any]:
"""只检索基线（GUIDELINES §7）：不调模型，检索第一名就当作已响应。"""
del ctx, job_type, job_id, task, prompt_id, version
frags = {m.group(1): m.group(2) for line in user.splitlines() if (m := FRAG.match(line.strip()))}
results, key = [], None
for line in user.splitlines():
if m := ITEM.match(line.strip()):
key = m.group(1)
elif (m := CANDS.match(line.strip())) and key:
first = m.group(1).split("、")[0]
results.append(
{"item": key, "verdict": "responded", "evidence": [{"sid": first, "quote": frags.get(first, "")}]}
)
return {"results": results}
def checklist(tender: str, run: str) -> dict[UUID, dict[str, Any]]:
"""读标评测输出的正式清单 → 查标的清单项（和检查快照的 payload 同样的字段）。"""
rows = json.loads((EVALS / "out" / "runs" / run / tender / "items.json").read_text())
@@ -94,6 +118,7 @@ async def main() -> None:
ap.add_argument("bid")
ap.add_argument("--checklist", default="fix3")
ap.add_argument("--name", default="base")
ap.add_argument("--baseline", action="store_true", help="只检索、不调模型的基线")
args = ap.parse_args()
d = OUT / args.bid
manifest = json.loads((d / "manifest.json").read_text())
@@ -102,7 +127,7 @@ async def main() -> None:
for n, pdf in enumerate(sorted(d.glob("v*.pdf")), start=1):
pages = text_layer.extract(f"{args.bid}-{pdf.stem}", pdf.read_bytes())
docs.append(match.BidDoc(uuid4(), n, pdf.stem, text_layer.body_sentences(pages)))
ai_service.run_json = cached_run_json # type: ignore[assignment]
ai_service.run_json = baseline_run_json if args.baseline else cached_run_json # type: ignore[assignment]
async def progress(_: float) -> None:
return None

183
evals/check/score.py Normal file
View File

@@ -0,0 +1,183 @@
"""查标评分：对照响应清单给一次离线检查打分（evals/check/GUIDELINES.md §7）。
cd server && ../.venv/bin/python ../evals/check/score.py <投标文件> [--run base] [--include-unreviewed]
- 清单项 → 标注单元：清单项的招标侧句子和单元的句子有重叠就对上（和读标评分同一口径）；
- 按单元去重：同一要求在清单里有几条，只要一条报出就算召回，几条都报不算误报；
- 问题类型按类型族比（D01 / D02 / D09 / D11 / D13 → unresponded，D03 / D07 → suspected_unmet，……）；
- 只算复核过的问题（决策 E1）；--include-unreviewed 时也算没复核的，报告里标明；
- 读标没抽到的问题单列「读标漏抽」：不算查标漏报，但计入端到端召回；
- 对不上任何标注单元的问题单列「无法对照」。
输出：evals/out/check/<投标文件>/runs/<运行>/score.json、score.md。
"""
from __future__ import annotations
import argparse
import json
import pathlib
import sys
from collections import Counter, defaultdict
from typing import Any
EVALS = pathlib.Path(__file__).resolve().parents[1]
sys.path.insert(0, str(EVALS))
import gold # noqa: E402
OUT = EVALS / "out" / "check"
FAMILY = {
"D01": {"unresponded"},
"D02": {"unresponded"},
"D09": {"unresponded"},
"D11": {"unresponded"},
"D13": {"unresponded"},
"D16": {"unresponded"},
"D03": {"suspected_unmet"},
"D07": {"suspected_unmet"},
"D08": {"suspected_unmet"},
"D05": {"negative_deviation", "suspected_unmet"},
"D06": {"negative_deviation"},
"D10": {"negative_deviation"},
"D12": {"qualification"},
"D14": {"qualification"},
"D15": {"qualification", "unresponded"},
}
FATAL_CATS = {"substantive", "qualification"}
def pct(a: int, b: int) -> str:
return f"{a}/{b}（{a / b:.0%}）" if b else "—"
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("bid")
ap.add_argument("--run", default="base")
ap.add_argument("--include-unreviewed", action="store_true")
a = ap.parse_args()
d = OUT / a.bid
manifest = json.loads((d / "manifest.json").read_text())
res = json.loads((d / "runs" / a.run / "results.json").read_text())
units, _ = gold.units(manifest["tender"])
sid_units: dict[str, set[str]] = defaultdict(set)
for u in units:
for s in u.sids:
sid_units[s].add(u.id)
per_unit: dict[str, list[dict[str, Any]]] = defaultdict(list)
unmapped: list[dict[str, Any]] = []
for it in res["items"]:
hit = {u for s in it["sentences"] for u in sid_units.get(s, ())}
for u in hit:
per_unit[u].append(it)
if not hit and it["finding"]:
unmapped.append(it)
rec: Counter[str] = Counter()
rec_total: Counter[str] = Counter()
sev_ok = sev_n = 0
misses: list[str] = []
extraction_miss: list[str] = []
fp: list[str] = []
fp_by_trap: Counter[str] = Counter()
ok_units = trap_units = 0
other_findings = 0
for uid, e in manifest["units"].items():
items = per_unit.get(uid, [])
found = [it for it in items if it["finding"]]
if e["expect"] == "defect":
if not e.get("reviewed") and not a.include_unreviewed:
continue
keys = [e["defect"], f"粒度:{e['granularity']}", "带标志词" if e.get("worded") else "不带标志词"]
fatal = e["cat"] in FATAL_CATS
if fatal:
keys.append("致命级")
for k in keys:
rec_total[k] += 1
rec_total["端到端"] += 1
if not items:
extraction_miss.append(f"{uid}（{e['defect']}，{e['cat']}）")
continue
rec_total["查标"] += 1
good = [it for it in found if it["finding"]["type"] in FAMILY.get(e["defect"], set())]
if good:
for k in [*keys, "端到端", "查标"]:
rec[k] += 1
if fatal:
sev_n += 1
sev_ok += any(it["finding"]["severity"] == "fatal" for it in good)
else:
got = ", ".join(sorted({it["verdict"] for it in items}))
misses.append(f"{uid}（{e['defect']} {e.get('variant', '')}）→ 查标结论：{got}")
elif e["expect"] == "ok":
ok_units += 1
if e.get("trap"):
trap_units += 1
if found:
fp.append(f"{uid}（{e['cat']}，{e['mode']}）：{found[0]['finding']['title'][:40]}")
if e.get("trap"):
fp_by_trap[e["trap"]] += 1
elif found:
other_findings += 1
findings = [it for it in res["items"] if it["finding"]]
lines = [
f"# 查标评分：{a.bid} · 运行 {a.run}",
"",
f"- 投标文件：{manifest['kind']}，风格 {manifest['style']}，{manifest['split']}；清单 {len(res['items'])} 条，报出问题 {len(findings)} 个",
f"- 金标准：{'含未复核的问题（试跑口径）' if a.include_unreviewed else '只算复核过的问题'}；复核方式见决策 E1",
f"- 模型：{res['summary']['usage']}",
"",
"## 召回",
"",
f"- 查标召回（读标抽到的要求上）：{pct(rec['查标'], rec_total['查标'])}",
f"- 端到端召回（含读标漏抽）：{pct(rec['端到端'], rec_total['端到端'])}",
f"- 致命级：{pct(rec['致命级'], rec_total['致命级'])}；其中级别也判成致命的 {pct(sev_ok, sev_n)}",
"",
"| 分类 | 召回 |",
"|---|---|",
*[f"| {k} | {pct(rec[k], rec_total[k])} |" for k in sorted(rec_total) if k not in ("端到端", "查标", "致命级")],
"",
"## 误报",
"",
f"- 规矩响应的要求 {ok_units} 条，报了问题的 {len(fp)} 条（每 100 条 {100 * len(fp) / ok_units:.1f}）"
if ok_units
else "- 没有规矩响应的要求",
f"- 干扰项 {trap_units} 条，误报 {dict(fp_by_trap)}",
f"- 要求本来不需要响应或无法判断（na / unknown）却报了问题：{other_findings}",
f"- 无法对照（对不上任何标注单元）的问题：{len(unmapped)}",
"",
f"## 漏报（{len(misses)}）",
"",
*[f"- {x}" for x in misses],
"",
f"## 读标漏抽（{len(extraction_miss)}）",
"",
*[f"- {x}" for x in extraction_miss],
"",
f"## 误报明细（{len(fp)}）",
"",
*[f"- {x}" for x in fp[:80]],
*(["- ……"] if len(fp) > 80 else []),
"",
f"## 规则问题（第 2 步才查，这里只列出）：{len(manifest['rule_defects'])} 个",
"",
]
out = d / "runs" / a.run
(out / "score.md").write_text("\n".join(lines))
summary = {
"recall": {k: [rec[k], rec_total[k]] for k in rec_total},
"fp": len(fp),
"ok_units": ok_units,
"fp_by_trap": dict(fp_by_trap),
"unmapped": len(unmapped),
"other_findings": other_findings,
"extraction_miss": len(extraction_miss),
"include_unreviewed": a.include_unreviewed,
}
(out / "score.json").write_text(json.dumps(summary, ensure_ascii=False, indent=1))
print("\n".join(lines[:20]))
if __name__ == "__main__":
main()

21
evals/check/vocab-dev.json Normal file
View File

@@ -0,0 +1,21 @@
{
"_comment": "开发集用的注入词库（GUIDELINES §3.4）。留出集用另一份，由另一个会话写好封存在 evals/out/check/_sealed/vocab-holdout.json，开发时不打开。check.match 提示词的例子不能和这里的短语相同。",
"D05": {"phrases": ["暂不支持{param}", "{param}需二次开发后支持", "{param}为可选配置"]},
"D06": {"deviations": ["负偏离", "部分满足", "偏离"]},
"D08": {"phrases": ["满足", "完全响应", "符合要求"]},
"D09": {"templates": ["{other}"]},
"D10": {"cells": ["未完全响应", "部分响应", ""]},
"P04": {"variants": ["drop_digit", "extra_shi", "decimal_shift"]},
"P06": {
"foreign_projects": [
{"name": "宿迁市人民医院智慧病房建设项目", "code": "JSZC-321300-HDZB-G2025-0318", "purchaser": "宿迁市人民医院"},
{"name": "盐城市大数据中心容灾备份系统采购项目", "code": "JSZC-320900-YCZB-G2025-0742", "purchaser": "盐城市大数据管理局"},
{"name": "泰州市第二中学智慧校园二期项目", "code": "JSZC-321200-TZJY-G2025-0096", "purchaser": "泰州市第二中学"}
]
},
"P07": {"blanks": ["__________", "年 月 日", "（ ）"]},
"P08": {"dates": ["2026年12月28日", "2027年11月20日", "2026年13月02日"]},
"P09": {"days": [60, 30, 45]},
"P10": {"phrases": ["本报价不含税，税费另行计算。", "付款方式双方另行协商。", "最终数量以实际到货为准。"]},
"P12": {"phrases": ["备选方案报价：{alt}元", "优惠报价：{alt}元", "方案二总价：{alt}元"]}
}