feat: 结构化输出按官方规范 + 两段式重试 + DeepSeek 复检开关

- 豆包: response_format.json_schema(strict);DeepSeek: 切 Responses API text.format.json_schema
  (官方 chat/completions 通道不支持 json_schema),不支持时自动降级 json_object([output] schema)
- 重试两段式:传输错误 cfg.retries 次;解析失败/空正文有独立 3 次专用重试,仍失败落 parse_fail
- 移除三票复核(decide_final/_verify_clean/verify_clean),DeepSeek 单次判定即终稿
- 新增 [run] deepseek_recheck 开关(默认 no=仅豆包初筛;yes=追加 DeepSeek 复检)
- GUI 不再覆盖 config 提示词;prompt 相对/前导斜杠路径按 exe 目录解析
- 报表去「票型/复核」列并同步说明;README/config.example.ini 同步
- 测试新增输出格式、重试、开关与提示词路径用例(50 passed)
This commit is contained in:
2026-09-03 11:44:10 +08:00
parent ae3a4e06f1
commit b413104587
12 changed files with 832 additions and 289 deletions
+3 -34
View File
@@ -1,7 +1,7 @@
# -*- coding: utf-8 -*-
"""pipeline 模块测试:排序、投票裁决、成本计算、缓存路径。"""
"""pipeline 模块测试:排序、成本计算、缓存路径。"""
from violation_detector.pipeline import (
_cache_path, _cat_num, deepseek_cost, decide_final, natural_key,
_cache_path, _cat_num, deepseek_cost, natural_key,
)
@@ -19,42 +19,11 @@ def test_natural_key_sort():
def test_cat_num():
assert _cat_num("12. 侵权 - 其他") == 12
assert _cat_num("1. 烟草") == 1
# 无编号的结论统一排到最后(无违规/违规不变动分类,不影响裁决
# 无编号的结论统一排到最后(无违规/违规不明不变动分类)
assert _cat_num("无违规") == 99
assert _cat_num("违规不明") == 99
def test_decide_final_unanimous_clean():
final, note = decide_final([R("无违规"), R("无违规"), R("无违规")])
assert final["category"] == "无违规"
assert "3票一致" in note
def test_decide_final_clean_majority_keeps_but_flags():
final, note = decide_final([R("无违规"), R("无违规"), R("12. 侵权 - 除人物外的其他侵权")])
assert final["category"] == "无违规"
assert "2:1" in note and "人工复核" in note
def test_decide_final_violation_majority_wins():
final, note = decide_final(
[R("无违规"), R("12. 侵权 - 除人物外的其他侵权"), R("12. 侵权 - 除人物外的其他侵权")])
assert final["category"].startswith("12.")
assert "人工复核" in note
def test_decide_final_three_way_split_takes_lowest_number():
final, note = decide_final([R("无违规"), R("13. 侵权 - 人物相关"), R("5. 负向敏感")])
assert final["category"].startswith("5.")
assert "从严" in note
def test_decide_final_two_violations_differ_majority_none():
# 两张违规票不同、一张无违规:违规方 2 票 > 1,并列取序号最小
final, _ = decide_final([R("无违规"), R("13. 侵权 - 人物相关"), R("12. 侵权 - 其他")])
assert final["category"].startswith("12.")
def test_deepseek_cost_known_usage():
# miss 407 + cached 1792 + completion 2282
results = {"a": R("无违规")}