"""出具校验的数字回指:叙述里写出来的每个量都得在指标集里找得到。 以前只认阿拉伯数字——"增长三成、客户两千五百家"在 strict 契约下判 formal, 同样的意思写成 2500 却会被 withheld。叙述层换一种写法就绕过了整套"不许现编 数字"的机制,而"出错自动改一次"的复核回路甚至会把模型往这条路上推。 中文数字难在两头:漏抽就是绕法,误抽就是满篇降档(一个问题、十分重要、从三个 方面看都不是数据)。所以这组用例两头都守。 """ from __future__ import annotations import pytest from app.engine import review as rv from app.engine.issuance import cn_to_number, decide_tier, extract_numbers, trace_numbers def values(text: str) -> list[tuple[str, float]]: return [(t.raw, round(t.value, 2)) for t in extract_numbers(text)] @pytest.mark.parametrize("text, expected", [ ("十二", 12), ("二十三", 23), ("一百零五", 105), ("两千五百", 2500), ("两千五", 2500), # 省略写法:千后面省掉了"百" ("一万二", 12000), ("三万零五百", 30500), ("一亿两千万", 1.2e8), ("三点五", 3.5), ("三点五万", 35000), ("三万亿", 3e12), ("万", 1e4), ("五一", None), ("三三两两", None), ("一一", None), ("十十", None), ("万一", None), # 不是一万一千 ("二〇二六", None), # 年份的逐位写法不是一个数 ]) def test_cn_to_number(text, expected): assert cn_to_number(text) == (None if expected is None else float(expected)) @pytest.mark.parametrize("text, expected", [ ("本周营收增长三成", [("三成", 30.0)]), ("客户两千五百家", [("两千五百", 2500.0)]), ("其中一半来自华东", [("一半", 50.0)]), ("翻了两倍", [("两倍", 2.0)]), ("约三分之一的门店", [("三分之一", 33.33)]), ("达到百分之三十", [("百分之三十", 30.0)]), ("关闭了两家门店,用时五天", [("两", 2.0), ("五", 5.0)]), ("新增两千余家", [("两千余", 2500.0)]), # 按区间 [2000, 3000] 回指 ("营收 3.5万,客户 2千家", [("3.5万", 35000.0), ("2千", 2000.0)]), ]) def test_quantities_written_in_chinese_are_extracted(text, expected): assert values(text) == expected @pytest.mark.parametrize("text", [ "这是一个积极的信号,一些门店一起统一调价,十分重要", "万一出问题千万不要慌,五花八门、千方百计、独一无二", "第三季度和三季度营收,周三开会", "十二月十五日,二〇二六年九月", "从三个方面看,主要原因之一", "几十家、数百家、上千家、十几家", "十一假期期间,下午三点给出三点建议", ]) def test_figurative_numerals_are_not_extracted(text): assert values(text) == [], "误抽了" def test_the_bypass_that_used_to_pass_as_formal(): """strict 契约下,指标集里没有的中文数字要让它 withheld,而不是盖 formal 章。""" narrative = "本周营收增长三成、客户两千五百家。" report = trace_numbers(narrative, [{"id": "revenue", "value": 1_234_567}]) assert [u["token"] for u in report.unmatched] == ["三成", "两千五百"] tier = decide_tier(missing_required=[], missing_expected=[], unmatched=report.unmatched, strict=True) assert tier == "withheld" def test_chinese_numerals_that_trace_back_pass(): """换成中文写法本身不是问题——回指得上就照常 formal。""" narrative = "本周营收增长三成、客户两千五百家,近一半来自华东。" metrics = [ {"id": "growth", "value": 0.2987}, # 比率:三成 = 30% ± 半成 {"id": "customers", "value": 2500}, {"id": "east_share", "value": 48.6}, # 百分数:一半 = 50% ± 5 ] report = trace_numbers(narrative, metrics) assert report.ok, report.unmatched assert decide_tier(missing_required=[], missing_expected=[], unmatched=[], strict=True) == "formal" def test_wan_suffix_tolerance_follows_written_precision(): """3.5万写到千位,按 ±500 回指——以前按 3.5 的一位小数给 ±0.05,只认 35000 整。""" assert trace_numbers("营收 3.5万", [{"id": "r", "value": 35123}]).ok assert not trace_numbers("营收 3.5万", [{"id": "r", "value": 36000}]).ok assert trace_numbers("营收 35万", [{"id": "r", "value": 351234}]).ok def test_ranges_trace_within_their_span(): assert trace_numbers("新增两千余家", [{"id": "n", "value": 2150}]).ok assert not trace_numbers("新增两千余家", [{"id": "n", "value": 3150}]).ok def test_review_may_rewrite_numbers_in_chinese_but_not_invent_them(): """复核重写时把 6 写成"六"不算新数字;多出来一个"五家"才算。""" assert rv.check_numbers("管理员有 6 个。", "管理员有六个。") == [] assert rv.check_numbers("管理员有 6 个。", "管理员有六个,商家有五家。") == ["五"]