合入复核修复(说法核对):结论句判定回到原来的词表(扩充的判断词只给裁判排先后)、表格正文不核、补因果与来源的口子、说法表按样本调整(96 句样本表)、连续按末尾方向核、复核认三种日志且只写说明、分析结论算可引用的来源、自动修复对认不出的取值照 claims 做、候选维度写明只认码值表、scan 加只按词认

This commit is contained in:
ThearchyHelios
2026-10-21 17:12:18 +08:00
12 changed files with 867 additions and 134 deletions
+129 -13
View File
@@ -43,7 +43,7 @@ def codes(found: list[dict]) -> list[str]:
@pytest.mark.parametrize("text, claim", [
("下降主要来自华东。", "source"), ("增长主要由线上贡献。", "source"), ("下降集中在华东、华北。", "source"),
("绝大部分来自线下。", "source"), ("华东拖累了整体。", "source"), ("线上带动了增长。", "source"),
("增长绝大部分来自线下。", "source"), ("华东拖累了整体。", "source"), ("线上带动了增长。", "source"),
("华东贡献最大。", "source"), ("下降主要是华东下滑造成的。", "source"),
("变化分散在各地区。", "dispersed"), ("各地区都在下降。", "dispersed"), ("各渠道普遍下滑。", "dispersed"),
("全面下降。", "dispersed"),
@@ -53,7 +53,7 @@ def codes(found: list[dict]) -> list[str]:
("销售额明显偏离往常。", "anomaly"), ("超出过去 12 周的常规波动范围。", "anomaly"),
("销售额明显高于正常水平。", "anomaly"), ("波动剧烈。", "anomaly"),
("本周创历史新高。", "record"), ("销售额创新高。", "record"), ("前所未有。", "record"),
("在常规波动范围内。", "normal"), ("属于正常波动。", "normal"), ("并不异常。", "normal"),
("在常规波动范围内。", "normal"), ("属于正常波动。", "normal"), ("这一降幅并不异常。", "normal"),
("线上销售额持续上升。", "trend"), ("呈上升趋势。", "trend"), ("一路走低。", "trend"), ("稳步增长。", "trend"),
("连续三周下降。", "trend"), ("连续下滑。", "trend"),
("没有明显的上升或下降趋势。", "no_trend"), ("上升趋势不明显。", "no_trend"), ("基本平稳。", "no_trend"),
@@ -138,13 +138,13 @@ def test_dafu_needs_a_baseline_outside_the_band_in_the_same_direction():
@pytest.mark.parametrize("text, claim", [
("并未出现明显异常。", "normal"), # 否定隔着「出现明显」
("本周没有出现异常。", "normal"),
("订单数并未出现明显异常。", "normal"), # 否定隔着「出现明显」
("本周销售额没有出现异常。", "normal"),
("不在正常范围内。", "anomaly"),
("不属于正常波动。", "anomaly"),
("没有显著高于华东。", "no_diff"),
("并非主要来自华东。", "source_not"),
("并不分散。", "source"),
("下降并非主要来自华东。", "source_not"),
("下降并不分散。", "source"),
("没有持续上升。", "no_trend"),
])
def test_a_negation_flips_to_the_opposite_class(text, claim):
@@ -196,8 +196,10 @@ def test_ascii_names_need_word_boundaries_but_chinese_neighbours_are_fine():
result["table"]["rows"][0][0] = "AB"
result["table"]["rows"][1][0] = "A"
assert said("下降主要来自AB店。", "A1", results={"A1": result}) == []
assert said("下降主要来自XAB。", "A1", results={"A1": result}) == [], "XAB 不是点名 AB,也没点名别的段"
assert codes(said("下降主要来自A。", "A1", results={"A1": result})) == [], "一个字的名字不认"
# XAB 不是点名 AB、一个字的名字不认:都当没点名哪一段,说的来源不是拆解里的段
[miss] = said("下降主要来自XAB。", "A1", results={"A1": result})
assert miss["code"] == ac.MISMATCH and "不是这份拆解里的段" in miss["message"]
assert codes(said("下降主要来自A。", "A1", results={"A1": result})) == [ac.MISMATCH]
def test_named_segments_are_matched_across_full_and_half_width():
@@ -254,6 +256,29 @@ def test_trends_check_the_direction_and_runs_check_the_length():
assert codes(said("线上销售额持续上升。", "A1", results={"A1": flat})) == [ac.MISMATCH]
def test_runs_follow_the_run_direction_and_the_number_of_periods_they_claim():
"""「连续 N 期上升」按趋势结论记的末尾连续同向核(run_direction、run_length):方向一致、期数不少于 N(没写 N 的至少
3 期),不看整体的趋势判定——判定语在方向相反时写「但最近已连续 N 期上升」。"""
rising = fake_result("trend", None, "rising", direction="up", values={"run_length": "4", "run_direction": "up"})
assert said("线上销售额已连续 4 周上升。", "A1", results={"A1": rising}) == []
[miss] = said("线上销售额持续增长,已连续 5 周上升。", "A1", results={"A1": rising})
assert miss["code"] == ac.MISMATCH and miss["claim"] == "trend" and "只有 4 期,不够说连续 5 期" in miss["message"]
# 整体下降、最近连续上升:「连续上升」对得上,「持续上升」对不上
turned = fake_result("trend", None, "falling", direction="down", values={"run_length": "4", "run_direction": "up"})
assert said("订单数最近已连续四周上升。", "A1", results={"A1": turned}) == []
assert codes(said("订单数已连续三周下降。", "A1", results={"A1": turned})) == [ac.MISMATCH]
assert codes(said("订单数持续上升。", "A1", results={"A1": turned})) == [ac.MISMATCH]
# 没有明显趋势、末尾连续下降:也能说连续
flat = fake_result("trend", None, "no_trend", direction=None, values={"run_length": "3", "run_direction": "down"})
assert said("订单数三连降。", "A1", results={"A1": flat}) == []
assert codes(said("订单数四连降。", "A1", results={"A1": flat})) == [ac.MISMATCH]
assert codes(said("订单数三连涨。", "A1", results={"A1": flat})) == [ac.MISMATCH]
# 没记 run_direction 的老结论按整体的趋势方向
old = fake_result("trend", None, "rising", direction="up", values={"run_length": "3"})
assert said("订单数连续上升。", "A1", results={"A1": old}) == []
assert codes(said("订单数连续下降。", "A1", results={"A1": old})) == [ac.MISMATCH]
def test_differences_check_the_verdict_and_the_order_of_the_two_groups():
differs = fake_result("differ", "rate", "differs", subjects=[["西南"], ["华东"]], direction="up")
assert said("西南的退款率显著高于华东。", "A1", results={"A1": differs}) == []
@@ -330,11 +355,65 @@ def test_a_cause_outside_the_decomposition_needs_its_own_evidence():
{"text": "由于促销结束导致下降。", "inputs": True},
{"text": "由于数据只到 10 月 16 日,本周不完整。"},
{"text": "由于样本少,差异不能下结论。"},
{"text": "以上是本周下降的主要原因。"}, # 原因只是指代:列出原因的那几句各自核
{"text": "下降的原因主要是:"},
])
def test_guesses_quotes_inputs_and_data_caveats_are_not_unsupported_causes(sentence):
assert ac.check_sentence(ac.Sentence(**sentence), {}) == []
@pytest.mark.parametrize("text, found", [
("下降主要由于促销结束。", [(ac.CAUSAL, "causal")]), # 「主要由」不吃掉「由于」
("华东下降主要由于促销结束。", [(ac.CAUSAL, "causal")]), # 原因在「由于」后面,前面的华东不算
("促销结束导致华东下降。", [(ac.CAUSAL, "causal")]), # 原因在「导致」前面
("华东受促销结束影响下降。", [(ac.CAUSAL, "causal")]), # 原因在「受…影响」中间
("下降主要由促销结束造成。", [(ac.MISMATCH, "source"), (ac.CAUSAL, "causal")]),
("下降主要来自线上渠道。", [(ac.MISMATCH, "source")]), # 说的来源不是拆解(按地区)里的段
("下降主要来自新客流失。", [(ac.MISMATCH, "source")]),
("华东下降主要来自线上渠道。", [(ac.MISMATCH, "source")]), # 「主要来自」看后面
("下降主要由华东造成。", []),
("华东是本周下降的主要原因。", []), # 原因点名的是段:按来源核
("下降的主要原因是华东门店客流减少。", []),
("华南是本周下降的主要原因。", [(ac.MISMATCH, "source")]),
("下降主要源于华东。", []),
("华东下滑,拖累了整体。", []), # 拖累看整个分句,没点名的按主要来源
])
def test_sources_and_causes_attached_to_a_decomposition(text, found):
"""挂了按地区拆解、判「主要来自华东」的结论(复核 R4):原因按因果词的那一边认;说的来源得是拆解里的段。"""
assert [(v["code"], v["claim"]) for v in said(text, "A1", results={"A1": fake_result()})] == found
def test_a_source_needs_no_name_when_the_clause_quotes_the_phrase_or_says_structure():
driver = fake_result()
phrase = ac.Sentence(text="下降主要来自以下地区:\x00\x00\x00\x00。", analyses=["A1"], phrases=[(11, 15, "A1")])
assert ac.check_sentence(phrase, {"A1": driver}) == []
elsewhere = ac.Sentence(text="\x00\x00\x00\x00,下降主要来自这一地区。", analyses=["A1"], phrases=[(0, 4, "A1")])
assert codes(ac.check_sentence(elsewhere, {"A1": driver})) == [ac.MISMATCH], "判定语在别的分句"
mix = fake_result(mode="ratio", verdict="mix_driven", subjects=[["华东"]])
assert said("退款率上升主要来自结构变化。", "A1", results={"A1": mix}) == []
rate = fake_result(mode="ratio", verdict="rate_driven", subjects=[["华东"]])
assert said("退款率上升主要来自各段自身的变化。", "A1", results={"A1": rate}) == []
assert [v["claim"] for v in said("退款率上升主要来自结构变化。", "A1", results={"A1": rate})] == ["source", "mix"]
def test_before_side_sources_look_back_to_the_nearest_named_clause():
driver = fake_result()
assert said("华东门店表现最弱,贡献最大。", "A1", results={"A1": driver}) == []
assert said("华南基本持平,华东贡献最大。", "A1", results={"A1": driver}) == [], "最近的分句点了名就不再往前找"
assert codes(said("新客贡献最大。", "A1", results={"A1": driver})) == [ac.MISMATCH]
@pytest.mark.parametrize("text, results, found", [
("下降主要由华东造成。", {}, [ac.UNCITED]), # 没有拆解可比:并进来源
("下降主要由华东造成。", {"A2": fake_result()}, [ac.UNCITED]), # 华东是目录里拆解的段:并进来源
("下降主要由促销结束造成。", {"A2": fake_result()}, [ac.UNCITED, ac.CAUSAL]),
("下降主要由促销结束造成。", {}, [ac.UNCITED]),
])
def test_without_a_decomposition_attached_a_cause_merges_into_the_source_only_when_it_could_be_one(text, results,
found):
assert codes(said(text, results=results)) == found
# --------------------------------------------------------------------------
# 文档:哪里不扫、怎么挂、版本
# --------------------------------------------------------------------------
@@ -413,6 +492,19 @@ def test_system_rendered_phrases_quotes_headings_and_code_are_not_scanned():
assert codes(ac.check_doc(doc, catalog())) == [ac.UNCITED]
def test_table_cells_written_by_the_writer_are_not_checked():
"""表格的格不核(表头、正文都不核):一格是一个值不是一句话,状态列的「异常」、说明列的「主要来自线上」不是判断
(judge.claim_priority 也不送表格的格)。表格外的段落、列表照核。"""
raw = ("本周销售额 [[m:gmv]]。[[see:m:gmv]]\n\n| 门店 | 状态 | 说明 |\n|---|---|---|\n| 甲店 | 异常 | 主要来自线上 |\n"
"| 乙店 | 正常 | 持续增长 |\n\n```\n异常订单激增\n```\n\n## 主要来自华东的下降\n\n- 华东拖累明显\n")
doc = compose_doc(raw, catalog(), analysis_claims="check")
tables = {u["id"] for b in doc["blocks"] if b["type"] == "table" for u in b["units"]}
assert len(tables) == 9, "表格切成了一格一格"
[v] = found(doc)
assert v["code"] == ac.UNCITED and v["text"] == "拖累" and v["unit"] not in tables
assert [v["text"] for v in found(verify_doc(doc, catalog()))] == ["拖累"]
def test_rendered_values_name_segments_but_are_not_judgement_words():
# [[v:…]] 渲染出来的「华东」能点名,渲染出来的「异常」不是写作者下的判断
doc = {"markdown": "", "blocks": [{"type": "paragraph", "units": [{"id": "u0", "kind": "claim", "cites": ["A1"],
@@ -438,6 +530,24 @@ def test_compose_and_verify_bind_the_sentence_to_its_see_marker():
assert [v["code"] for v in verify_doc(tampered, cat)["violations"]] == [ac.UNCITED]
@pytest.mark.parametrize("raw, want", [
("销售额 [[m:gmv]],下降主要来自线上渠道。[[see:A1]]", [(ac.MISMATCH, "主要来自")]),
("销售额 [[m:gmv]],下降主要来自新客流失。[[see:A1]]", [(ac.MISMATCH, "主要来自")]),
("销售额 [[m:gmv]],下降主要由于促销结束。[[see:A1]]", [(ac.CAUSAL, "由于")]),
("销售额 [[m:gmv]],下降主要由促销结束造成。[[see:A1]]", [(ac.MISMATCH, "主要由"), (ac.CAUSAL, "造成")]),
("销售额 [[m:gmv]],下降主要来自华东。[[see:A1]]", []),
("销售额 [[m:gmv]],下降主要来自 [[v:A1.r0.region]]。[[see:A1]]", []),
("销售额 [[m:gmv]],下降主要来自以下地区:[[a:A1]]。", []),
("[[a:A1]],下降主要来自这一地区。", [(ac.MISMATCH, "主要来自")]),
])
def test_strict_documents_close_the_source_and_cause_loopholes(raw, want):
"""复核 R4 的几个口子(strict 下挂一份按地区拆解、判「主要来自华东」的 A1):经组装和复核都核出来。"""
cat = catalog(A1=fake_result())
doc = compose_doc(raw, cat, analysis_claims="strict")
assert [(v["code"], v["text"]) for v in found(doc)] == want
assert [(v["code"], v["text"]) for v in found(verify_doc(doc, cat))] == want
def test_inputs_and_quotes_in_the_same_sentence_support_a_cause():
doc = compose_doc("由于本周有 [[i:promo]] 活动,销售额 [[m:gmv]]。[[see:m:gmv]]", catalog(), analysis_claims="check")
assert found(doc) == []
@@ -454,12 +564,18 @@ def test_the_prompt_rules_depend_on_whether_there_are_conclusions():
assert "[[a:A1]]" in ac.ANALYSIS_RULES_PROMPT
def test_judgement_words_make_a_sentence_a_claim_for_the_judge():
def test_judgement_words_rank_claims_for_the_judge_but_do_not_make_sentences_claims():
"""判断词只管裁判送判的先后(judge._JUDGEMENT):一句话算不算结论仍按原来的方向词、因果词(evidence._DIRECTIONAL),
结论句策略要求挂依据、出具缺口照它数——扩了就改了已发布版本的档位。"""
from app.engine.judge import claim_priority
doc = compose_doc("华东异常。\n\n下面主要看三个方面。\n\n建议持续关注。\n\n近 8 周的趋势图如下:", catalog())
doc = compose_doc("华东异常。\n\n华东异常,退款 12 单。\n\n下面是主要渠道的表现。\n\n促销期间活动效果显著。\n\n"
"整体运行平稳,未见异常。\n\n建议持续关注。\n\n近 8 周的趋势图如下:\n\n华东退款 12 单。", catalog())
kinds = {doc["markdown"][slice(*u["span"])]: (u["kind"], claim_priority(u, b, doc["markdown"]))
for b, u in evidence.iter_units(doc)}
assert kinds["华东异常。"] == ("claim", 2)
assert kinds["下面主要看三个方面。"][0] == "connective" and kinds["建议持续关注。"][0] == "connective"
assert kinds["近 8 周的趋势图如下:"][0] == "connective", "引出图表的话不算判断"
for text in ("华东异常。", "下面是主要渠道的表现。", "促销期间活动效果显著。", "整体运行平稳,未见异常。", "建议持续关注。",
"近 8 周的趋势图如下:"):
assert kinds[text] == ("connective", None), text
assert doc["stats"]["uncited_claims"] == 2, "只有两句带数字的算结论"
assert kinds["华东异常,退款 12 单。"] == ("claim", 2), "结论句里的判断词让裁判先判它"
assert kinds["华东退款 12 单。"] == ("claim", 3)
@@ -0,0 +1,154 @@
"""说法表的样本表(engine/analysis_claims,细案 01 §9.1):常见的周报写法逐句过 scan,每句期望认出的类(按出现的先后,
否定翻过之后的)或「不核」(空)。以后改说法表都过这张表:误报(正常的句子被打回)让报告多重写一次、strict 下降档,
漏报只是少核一处,宁可漏报不误报。
前半是第 1 波复核时的 70 句(复核第 4 节),后半是这次调整顺带看清楚的写法。和复核标记不一样的只有一句:「退款主要集中在
生鲜品类」复核标的是判断,可它说的是退款的构成、没说变化,按「来源、集中只在说了变化时才核」不核。
"""
from __future__ import annotations
import pytest
from app.engine import analysis_claims as ac
#: 不核
NO: tuple[str, ...] = ()
SAMPLES: list[tuple[str, tuple[str, ...]]] = [
# ---- 陈述数字和变化:不核 ----
("本周销售额 12.3 万元,环比下降 5.2%。", NO),
("整体平稳,各品类变化不大。", NO),
("退款率略有下降。", NO),
("订单数明显回升,客单价基本持平。", NO), # 「明显」只是程度,不收「明显 + 涨跌」
("环比基本持平。", NO),
("本周数据只到 10 月 16 日,周数据不完整。", NO),
("客单价相当于上周的 1.1 倍。", NO),
("订单量与上周相当。", NO), # 比上期:说的是持平,不是两组差异
# ---- 判断:核 ----
("销售额同比增长 8.1%,主要是因为国庆假期客流增加。", ("causal",)),
("受节假日影响,线下门店客流明显回升。", ("causal",)),
("华东销售额大幅增长 32%。", ("anomaly",)),
("线上渠道拉动整体增长。", ("source",)),
("华东下滑拖累了整体表现。", ("source",)),
("各地区普遍上涨。", ("dispersed",)),
("已连续三周下降。", ("trend",)),
("西南地区退款率显著高于其他地区。", ("differ",)),
("两个渠道的转化率相当。", ("no_diff",)),
("促销效果显著,带动订单量增长。", ("source",)),
("由于华东仓库搬迁,发货延迟导致退款增加。", ("causal", "causal")),
("线上销售额持续增长,已连续 5 周上升。", ("trend", "trend")),
("退款金额骤增,需重点关注。", ("anomaly",)), # 「需」在后一个分句,不管前面的判断
("新客贡献最大。", ("source",)),
("这主要是由于华东仓库搬迁。", ("causal",)),
("整体来看,各渠道均有所下降。", ("dispersed",)),
("销量下降主要由促销结束造成。", ("source", "causal")),
("下降主要来自客流减少。", ("source",)),
("退款主要集中在生鲜品类。", NO), # 复核标的是判断;说的是构成、没说变化,不核
# ---- 业务上的「异常」、数据质量:不核 ----
("本周异常订单 35 单,已全部处理。", NO),
("部分门店数据异常,已提交核查。", NO),
("部分订单异常,已退款。", NO),
("物流异常的订单已全部退款。", NO),
("支付异常的订单已全部退款。", NO),
("门店营业异常 2 家。", NO),
("整体运行平稳,未见异常。", NO), # 主语不是指标、波动
("系统运行正常,无异常。", NO),
# ---- 说的是构成,不是变化从哪来:不核 ----
("本周订单主要来自线上渠道,占 65%。", NO),
("销售额集中在周末两天。", NO),
("华东门店集中在上海、杭州。", NO),
("客户分散在各个城市。", NO),
("线上渠道占主导。", NO),
# ---- 建议、计划、预测:不核 ----
("下一步将全面提升门店服务水平。", NO),
("建议加大投放,拉动新客增长。", NO), # 「建议」管到后面的分句
("通过会员活动带动复购。", NO), # 带动的不是变化
("下周将持续提升转化率。", NO),
("建议持续降低退款率。", NO),
("计划逐步提高线上占比。", NO),
("下一步将大幅提升配送效率。", NO),
("下周重点主要是提升转化率。", NO),
("避免缺货导致客诉。", NO),
("预计下周将继续保持增长态势。", NO),
# ---- 别的误报:不核 ----
("订单来源于线上渠道。", NO), # 「来源于」不是「源于」
("数据来源于业务系统,金额含税。", NO),
("本周数据源表结构变化,部分字段暂缺。", NO), # 表结构不是结构变化
# ---- 漏报,补上:核 ----
("华东是本周下降的主要原因。", ("causal",)), # 原因点名的是拆解的段时按来源核
("近七成的降幅来自华东。", ("source",)),
("超过一半的下降来自华东。", ("source",)),
("增长主要靠线上。", ("source",)),
("降幅主要在华东。", ("source",)),
("下降主要体现在华东。", ("source",)),
("增长以线上为主。", ("source",)),
("销售额创年内新高。", ("record",)),
("销售额创近一年新高。", ("record",)),
("整体趋势向好。", ("trend",)),
("本周销售额波动较大。", ("anomaly",)),
("订单量三连降。", ("trend",)),
("退款率显著下降。", ("anomaly",)),
("销售额远高于往常。", ("anomaly",)),
("本周销售额出现异动。", ("anomaly",)),
("华东和华南一涨一跌。", ("offset",)),
# ---- 这次调整顺带看清楚的写法 ----
("订单大幅减少。", ("anomaly",)), # 「订单」只在「订单异常」里是业务状态
("不少门店大幅下滑。", ("anomaly",)),
("订单异常下降。", ("anomaly",)), # 「异常」后面跟着涨跌:说的是指标
("华东线上订单出现异常的下降。", ("anomaly",)),
("异常的门店已全部整改。", NO),
("两个渠道的转化率与上周相当。", NO),
("华东和华南的退款率水平相当。", ("no_diff",)),
("销售额并未出现明显异常。", ("normal",)),
("本周没有出现异常。", NO),
("退款率并未显著下降。", ("normal",)),
("下降主要来自华东,建议加大华东投放。", ("source",)), # 「建议」在后一个分句
("销售额\x00\x00,环比\x00\x00,主要来自华东。", ("source",)), # 引用渲染的数涂成 \x00,「环比」说的是变化
("订单主要来自线上渠道,占比较上周提升 5 个百分点。", NO), # 变化在后一个分句:前面说的是构成
("新增客户主要来自线上渠道。", NO),
("增量主要来自线上。", ("source",)),
("下降主要源于华东。", ("source",)),
("订单以线上为主。", NO),
("线上渠道表现突出,贡献最大。", ("source",)),
("下降的主要原因是华东门店客流减少。", ("causal",)),
("受促销结束影响,华东销售额下降。", ("causal",)),
("近 8 周销售额持续下降,已连续三周下降。", ("trend", "trend")),
("线上销售额十二连涨。", ("trend",)),
("整体走势向下。", ("trend",)),
("销售额创下了近一年来新高。", ("record",)),
("创新能力最高的门店是华东店。", NO),
("预计下周销售额将创新高。", NO),
]
@pytest.mark.parametrize("text, expected", SAMPLES, ids=[t for t, _ in SAMPLES])
def test_the_sample_table(text, expected):
found = ac.scan(text)
assert tuple(h.claim for h in found) == expected, [(h.claim, h.word) for h in found]
@pytest.mark.parametrize("text, expected", SAMPLES, ids=[t for t, _ in SAMPLES])
def test_a_sentence_is_reported_without_conclusions_exactly_when_it_carries_a_judgement(text, expected):
"""没有分析结论可挂时:认出判断的句子至少报一处(没挂、原因没依据),不核的一处都不报。"""
assert bool(ac.check_sentence(ac.Sentence(text=text), {})) == bool(expected)
def test_the_sample_table_is_not_shorter_than_the_review():
assert len(SAMPLES) >= 70 and len({t for t, _ in SAMPLES}) == len(SAMPLES)
@pytest.mark.parametrize("text, count", [("已连续三周下降。", 3), ("线上销售额持续增长,已连续 5 周上升。", 5),
("线上销售额十二连涨。", 12), ("连续两个月下滑。", 2), ("连续多周上升。", None),
("订单量三连降。", 3), ("连续10周上升。", 10)])
def test_runs_carry_the_number_of_periods_they_claim(text, count):
[run] = [h for h in ac.scan(text) if h.run]
assert run.count == count
@pytest.mark.parametrize("label, claims", [("主要来自华东的销售额", ["source"]), ("集中在线上的订单", ["source"]),
("建议持续降低的退款率", ["trend"]), ("部分订单异常", ["anomaly"]),
("销售额", [])])
def test_bare_scanning_finds_judgement_words_in_a_name(label, claims):
"""认叫法里夹带的判断词(bare=True)只按词认:叫法是一个名字不是一句话,构成、建议计划、业务状态这些整句话的规矩不用。"""
assert [h.claim for h in ac.scan(label, bare=True)] == claims
assert ac.scan(label) == []
@@ -95,6 +95,31 @@ def test_the_autofix_writes_it_in():
assert write["data"]["config"]["analysis_claims"] == "strict"
def test_an_unrecognized_value_is_left_to_its_own_fix_like_claims():
"""analysis_claims 写了认不出的值:门禁 G3 跳过、交给校验报,自动修复的 G3 那一条也不碰它(以前顺手改成了 strict),
由校验那条问题自己的修复管——照 claims 的做法:受管级别改成 strict,别的级别三档要人选。"""
graph = traceable({**STRICT, "numbers": "off", "analysis_claims": "loose"})
spec = GraphSpec.model_validate(graph)
fixes = {f["id"]: f for f in plan_fixes(spec, publish_issues(spec, level="governed"), level="governed")}
assert fixes["governed.report_policy:write"]["preview"] == {"field": "numbers", "before": "off", "after": "strict"}
bad = fixes["report.analysis_claims_invalid:write"]
assert bad["kind"] == "auto" and bad["preview"] == {"field": "analysis_claims", "before": "loose", "after": "strict"}
out = apply_fixes(graph, list(fixes), level="governed")
write = next(n for n in out["graph"]["nodes"] if n["id"] == "write")["data"]["config"]
assert out["ok"] is True and write["analysis_claims"] == "strict" and write["numbers"] == "strict"
published = traceable({"analysis_claims": "loose"})
spec = GraphSpec.model_validate(published)
fid = "report.analysis_claims_invalid:write"
[fix] = [f for f in plan_fixes(spec, publish_issues(spec, level="published"), level="published") if f["id"] == fid]
assert fix["kind"] == "choice" and [o["value"] for o in fix["options"]] == ["off", "check", "strict"]
assert [o["label"] for o in fix["options"]] == [option_label("analysis_claims", v) for v in ("off", "check", "strict")]
assert "default" not in fix
out = apply_fixes(published, [fid], {fid: "check"}, level="published")
assert out["applied"] == [fid] and next(n for n in out["graph"]["nodes"] if n["id"] == "write")[
"data"]["config"]["analysis_claims"] == "check"
def test_a_governed_upgrade_writes_the_strict_setting():
from app.engine.upgrade import _policy
@@ -265,6 +265,61 @@ async def test_the_scope_is_written_only_when_there_are_conclusions(monkeypatch,
assert "analysis_scope" not in doc
# --------------------------------------------------------------------------
# 上游的可引用来源认分析结论(00-总方案 D10)
# --------------------------------------------------------------------------
def card_entry(alias: str) -> dict:
"""分析卡做的结论(台账条目带 item、不带 call_id)。"""
return {**a_entry(alias, fake_result()), "item": "drop", "node_id": "card"}
def agent_entry(alias: str) -> dict:
"""Agent 的分析工具现场做的结论(带 call_id)。"""
return {**a_entry(alias, fake_result()), "call_id": "call-1", "node_id": "ask"}
@pytest.mark.parametrize("cells", [True, False])
def test_conclusions_from_an_analysis_card_are_citable_evidence(cells):
"""「证据来自」只选了分析卡:分析卡的结论总能引用(判定语、具名的数),不报「上游没有可引用的来源」;受管级别的正式
出具没开「单元格引用」时也不报「只有查询结果」。"""
assert report_mod._no_evidence({"A1": card_entry("A1")}, cells) is None
assert report_mod._no_evidence({"Q1": {"kind": "query"}, "A1": card_entry("A1")}, cells) is None
def test_conclusions_from_an_agent_follow_the_cells_rule():
catalog = {"Q1": {"kind": "query"}, "A1": agent_entry("A1")}
assert report_mod._no_evidence(catalog, True) is None
warning = report_mod._no_evidence(catalog, False)
assert warning and "只有查询结果和 Agent 现场做的分析" in warning and "无法引用单元格和这些分析结论" in warning
plain = report_mod._no_evidence({"Q1": {"kind": "query"}}, False)
assert plain and "只有查询结果,没有口径卡指标" in plain and "Agent" not in plain
def test_nothing_citable_names_analysis_conclusions_too():
warning = report_mod._no_evidence({}, True)
assert warning and all(kind in warning for kind in ("口径卡指标", "查询结果", "知识库检索", "运行输入", "分析结论"))
# --------------------------------------------------------------------------
# 判断词不改一句话算不算结论:已发布版本的档位不变
# --------------------------------------------------------------------------
@pytest.mark.parametrize("extra", ["整体运行平稳,未见异常。", "下面是主要渠道的表现。", "促销期间活动效果显著。"])
async def test_judgement_words_do_not_turn_connective_sentences_into_uncited_claims(monkeypatch, extra):
"""已发布的受管版本(结论句要挂依据、出口带契约):没有数字、没有引用的话照旧是连接性的话,不因为带了「主要」「异常」
「显著」就算没挂依据的结论、记缺口降档(判断词只给裁判排送判的先后,judge._JUDGEMENT)。"""
graph = weekly(numbers="strict", on_violation="flag", claims="require_citation")
graph["nodes"][-1]["data"]["config"]["contract"] = {
"report_from": "write", "metrics_from": ["caliber"], "required": ["gmv"], "strict": True}
script(monkeypatch, f"{PLAIN}\n\n{extra}")
issuance = (await finish(graph)).output["_issuance"]
assert issuance["tier"] == "formal" and issuance["gaps"] == []
assert issuance["claims"] == {"policy": "require_citation", "uncited_claims": 0, "uncited": []}
# --------------------------------------------------------------------------
# 取值写错
# --------------------------------------------------------------------------
+94
View File
@@ -476,3 +476,97 @@ def test_a_rewrite_counts_by_the_final_judgement():
judge={"counts": {"supported": 2, "contradicted": 1, "insufficient": 0}})]
assert rv.scan(good, OK_OUTPUT) == [] and rv.notes(good) == ["报告「写周报」有 1 句结论裁判认为证据不支持,改写后都有了依据"]
assert [s.kind for s in rv.scan(bad, OK_OUTPUT)] == ["report_rewrite"] and rv.notes(bad) == []
# --------------------------------------------------------------------------
# 分析能力一轮的信号:只写说明、不自动重来(00-总方案 §3.7)
# --------------------------------------------------------------------------
#: check 下判断词重写一次后仍没挂对分析结论:报告照常产出,标在报告里
CLAIMS_LEFT = "报告有 1 处分析说法未通过核对(1 处判断没挂分析结论:「主要来自」),已要求模型重写 1 次仍未改掉:…这几处已标注"
REFUSED = "「周报分析」的波动基线没有做成:历史只有 5 期,至少要 8 期"
SKIPPED = "指标「销售额环比」用到的两次查询没做成可比性检查:探查超时。指标照常计算,两期能不能直接比请自行核对"
def test_analysis_signals_are_registered_as_gaps_that_only_need_a_note():
signals = rv.scan([warn(CLAIMS_LEFT, "report_analysis_claims"), warn(REFUSED, "analysis_refused"),
warn(SKIPPED, "metric_comparability_skipped")], OK_OUTPUT)
assert [(s.kind, s.severity) for s in signals] == [
("report_analysis_claims", rv.DEGRADED), ("analysis_refused", rv.DEGRADED),
("metric_comparability_skipped", rv.DEGRADED)]
assert {s.kind for s in signals} == rv.NOTE_ONLY
assert not rv.should_retry(signals) and not rv.NOTE_ONLY & (rv.RETRYABLE | rv.REBUILD)
@pytest.mark.asyncio
async def test_a_run_with_only_note_only_signals_is_annotated_without_calling_the_model():
model = FakeModel(AssertionError("不该叫模型"))
signals = rv.scan([warn(CLAIMS_LEFT, "report_analysis_claims"), warn(REFUSED, "analysis_refused")], OK_OUTPUT)
result = await rv.review(model, question="q", answer="答案", signals=signals, locked=True)
assert model.calls == 0
assert result.verdict == "annotated" and result.retry is False and result.severity == rv.DEGRADED
assert result.note == f"{CLAIMS_LEFT};{REFUSED}" and result.answer is None
# 和别的信号一起出现时照常叫模型
mixed = [*signals, rv.Signal("tool_error", "工具 x 调用失败", rv.BROKEN)]
model = FakeModel({"note": "取数失败了一次。", "answer": "", "retry": True})
assert (await rv.review(model, question="q", answer="答案", signals=mixed)).note == "取数失败了一次。"
assert model.calls == 1
def _left_after_rewrite(mode: str, violations: list[dict], count: int = 2) -> list[dict]:
return [
ev(EventType.NODE_STARTED, node_id="write", label="写周报"),
ev(EventType.LOG, node_id="write", level="warn", code="report_repair", count=count, attempt=1,
message=f"报告有 {count} 处未通过核对(「12」、「主要来自」),已要求模型重写(第 1 次)"),
ev(EventType.REPORT_CHECKED, node_id="write", ok=not violations, violations=violations, repairs=1,
analysis_claims=mode),
]
CLAIM = {"code": "analysis_claim_uncited", "text": "主要来自", "message": "「主要来自」是分析判断"}
def test_check_mode_leftovers_do_not_make_the_rewrite_count_as_failed():
"""check 下重写一次后仍有的说法违规是标注(同缺失断言,各有自己的信号):「改过一次」算改好了,不再另报 report_repair;
复核被叫起来时照实说剩下的是哪几处。"""
rows = [*_left_after_rewrite("check", [CLAIM]), warn(CLAIMS_LEFT, "report_analysis_claims")]
assert [s.kind for s in rv.scan(rows, OK_OUTPUT)] == ["report_analysis_claims"]
assert rv.notes(rows) == ["报告「写周报」首稿有 2 处没过核对,重写后只剩已标出的 1 处分析说法"]
# 一处都没改掉:不说「改好了」,剩下的由它自己的信号说
assert rv.notes(_left_after_rewrite("check", [CLAIM], count=1)) == []
def test_absence_claim_leftovers_are_annotations_too():
absent = {"code": "absence_claim", "text": "没有查到", "message": "缺失断言"}
rows = _left_after_rewrite("check", [absent, CLAIM], count=3)
assert rv.scan(rows, OK_OUTPUT) == []
assert rv.notes(rows) == ["报告「写周报」首稿有 3 处没过核对,重写后只剩已标出的 1 处缺失断言、1 处分析说法"]
@pytest.mark.parametrize("mode, violations", [
("strict", [CLAIM]), # strict 下是出具缺口,不是标注
("check", [CLAIM, {"code": "uncited_number", "text": "12"}]), # 还有别的违规
("off", [CLAIM]),
("check", [CLAIM] * 20), # 带满了 20 条:后面可能还有
])
def test_other_leftovers_still_count_as_a_failed_rewrite(mode, violations):
rows = _left_after_rewrite(mode, violations)
assert [s.kind for s in rv.scan(rows, OK_OUTPUT)] == ["report_repair"] and rv.notes(rows) == []
@pytest.mark.asyncio
async def test_review_endpoint_does_not_call_the_model_for_a_check_mode_leftover(client, monkeypatch):
model = FakeModel(AssertionError("不该叫模型"))
async def fake_model(*a, **k):
return model, "fake"
monkeypatch.setattr("app.api.copilot.get_chat_model", fake_model)
output = {"周报": "周报:销售额 45,678.5元,下降主要来自华东。",
"_evidence": {"report_node": "write", "doc_artifact": "a" * 64, "fields": ["周报"]}}
run_id = await _make_run(client, output, [*_left_after_rewrite("check", [CLAIM]),
warn(CLAIMS_LEFT, "report_analysis_claims")])
body = (await client.post("/api/copilot/review", json={"run_id": run_id, "question": "q"})).json()
assert model.calls == 0
assert body["verdict"] == "annotated" and body["note"] == CLAIMS_LEFT and body["retry_instruction"] is None
assert [s["kind"] for s in body["signals"]] == ["report_analysis_claims"]