报告里照抄查询结果中文本格的名字(「冲压5线」「KX-205 线」)时,名字里的数字不再算裸数字:目录里每次查询的快照中、文本格里带数字的值(至少一个数字、至少一个不是数字的字,本身不是数也不是带单位的数),在正文里一字不差地出现、前后不紧挨着数字的那几段,查裸数字时涂白。组装和出口复核用同一条规则;只认快照里真有的值,不做模糊匹配,同一个数单独写出来照样要出处,数值格不受影响

This commit is contained in:
ThearchyHelios
2026-10-19 17:06:08 +08:00
parent 8f6a90c8c2
commit baf93ddfed
2 changed files with 123 additions and 3 deletions
@@ -0,0 +1,73 @@
"""查询结果里文本格的名字(「冲压5线」「KX-205 线」)照抄进报告时,名字里的数字不算裸数字。
部署上撞过:查询结果的产线名里带数字,写作者照着写「冲压5线的产量 [[v:…]]」,名字里的那个 5 被当成没有出处的数字打回
重写,重写还是同一个名字。现在:写作目录的查询结果里、文本格的值(一字不差的子串;值里至少有一个不是数字的字,本身
不是数、也不是带单位的数)出现在正文里时,那一段不查裸数字。只认快照里真有的值,不做模糊匹配:同一个数单独写出来照样
要出处;数值格不受影响。数据是虚构的。
"""
from __future__ import annotations
import pytest
from app.core import artifact_store
from app.engine.evidence import build_catalog, compose_doc, verify_doc
SNAP = {
"columns": ["产线", "产量", "备注"],
"column_types": {"产线": "text", "产量": "number", "备注": "text"},
"rows": [
["冲压5线", 1520, "120台"], # 带单位的数:不算名字
["KX-205 线", 830, "45%"],
["焊装 2 线", 64, "3.5万元"],
],
"row_count": 3, "truncated": False, "elapsed_ms": 2,
"sql": "SELECT line AS 产线, SUM(qty) AS 产量, MAX(memo) AS 备注 FROM output GROUP BY line",
"source": "plant",
}
def _store(snap: dict) -> str:
text = artifact_store.canonical_json(snap)
artifact = artifact_store.content_hash(text)
path = artifact_store._path_of(artifact)
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(text, encoding="utf-8")
return artifact
@pytest.fixture
def catalog():
art = _store(SNAP)
ledger = [{"kind": "query", "node_id": "fetch", "exec": 1, "artifact": art, "via": "b" * 64, "call_id": "call_1",
"tool": "db_query__plant", "source": "plant", "columns": SNAP["columns"], "rows": 3, "truncated": False}]
return build_catalog(nodes={}, ledger=ledger, inputs={})
def _uncited(doc) -> list[str]:
return [v["text"] for v in doc["violations"] if v["code"] == "uncited_number"]
def test_a_number_inside_a_name_from_the_results_is_allowed(catalog):
text = "冲压5线的产量是 [[v:Q1.r0.产量]],KX-205 线是 [[v:Q1.r1.产量]],焊装 2 线是 [[v:Q1.r2.产量]]。[[see:Q1]]"
doc = compose_doc(text, catalog)
assert _uncited(doc) == [], doc["violations"]
assert verify_doc(doc, catalog)["ok"], "出口复核按同一条规则:组装时放过的,复核也放过"
def test_the_same_number_on_its_own_still_needs_a_citation(catalog):
doc = compose_doc("冲压5线有 5 个班组,产量 [[v:Q1.r0.产量]]。[[see:Q1]]", catalog)
assert _uncited(doc) == ["5"]
# 名字只认一字不差的:快照里没有「冲压15线」,15 照样要出处;名字前后紧挨着数字的也不算那个名字
assert _uncited(compose_doc("冲压15线的产量 [[v:Q1.r0.产量]]。", catalog)) == ["15"]
assert _uncited(compose_doc("KX-2050 线的产量 [[v:Q1.r1.产量]]。", catalog)) == ["2050"]
assert _uncited(compose_doc("焊装 2 线以外还有 2 线。", catalog)) == ["2"]
def test_numbers_and_numbers_with_units_in_text_cells_are_not_names(catalog):
doc = compose_doc("备注里写着 120台、45%、3.5万元。", catalog)
assert _uncited(doc) == ["120", "45%", "3.5万"], doc["violations"]
def test_numeric_cells_are_unaffected(catalog):
# 产量格的 1520 是数值格:照抄成裸数字照样违规
assert _uncited(compose_doc("冲压5线的产量是 1520。", catalog)) == ["1520"]