张胜宇
e239eb778b
1) 客服 Agent 四份交付文档 + 构建脚手架：品牌由包装占位 XX科技 / 旧名 南方财富 统一为南方基金（热线 400-889-8899 / 官网 nffund.com），系统名改为「智能服务系统」； 同步追加 §0.4 修订记录行，工程记录行保留原占位字面以支撑硬编码扫描验收。 2) 开发文档：清理 28 份已作废/残留文档（14 份移出归档 + 14 份仓库副本）， 新增《文档规整方案与开发前待决事项-2026-09-17》。 3) 客服agent 四份交付文档首次纳入本分支。
109 lines
4.6 KiB
Python
109 lines
4.6 KiB
Python
"""知识块粒度选择的单元测试。
|
||
|
||
本文件当前只覆盖**检索层**（`KnowledgeSearchService`）的粒度归并。原先另有 4 条用例覆盖
|
||
`CustomerServiceAgent._prefer_section`（「整节块」与「行级子块」之间的取舍），已随客服
|
||
Agent 模块一并移除——重建客服 Agent 时须把该判据连同用例一起带回来。
|
||
|
||
检索层用例的背景是实测的三次翻车，每条判据都对应其中一次：
|
||
|
||
1. 客户问「起投多少」和「风险高吗」时命中同一块（整个产品小节），拿到**完全相同**的
|
||
整节内容，看起来像客服没听懂问题——所以把表格行拆成了行级子块。
|
||
2. 拆细之后「介绍一下」又被某一行抢答（返回"产品期限 90天封闭期"）——所以要能换回整节。
|
||
3. 两次判据写错：用"含连字符"认子块时，整节块自己的编号 PROD-901 被误判成子块；
|
||
用"不含两位数字后缀"认整节块时，FAQ 块全被误判成整节块、把正确答案挤出了 top1。
|
||
|
||
第 4 次翻车（2026-09-15）与"同节兄弟子块互相打平"有关：`doc_id` 去重挡不住
|
||
`POL-AST-009-07` 与 `POL-AST-009-12` 这种**同父不同子**，实测它们把「风险评估问卷怎么评分」
|
||
的 top1/次优差压到 0.002 → 客服判并列转人工。
|
||
"""
|
||
|
||
from app.service.knowledge_search_service import KnowledgeHit, KnowledgeSearchService
|
||
|
||
|
||
def hit(doc_id: str, score: float, content: str = "正文") -> KnowledgeHit:
|
||
return KnowledgeHit(doc_id=doc_id, title=doc_id, content=content, score=score)
|
||
|
||
|
||
def test_parent_of_recognises_row_blocks() -> None:
|
||
assert KnowledgeSearchService._parent_of("PROD-007-04") == "PROD-007"
|
||
|
||
|
||
def test_parent_of_rejects_section_blocks() -> None:
|
||
"""整节块的编号本身就含连字符（PROD-901），不能被当成子块。"""
|
||
assert KnowledgeSearchService._parent_of("PROD-901") is None
|
||
assert KnowledgeSearchService._parent_of("FAQ-0016") is None
|
||
assert KnowledgeSearchService._parent_of("HNW-003") is None
|
||
|
||
|
||
# --- 同节兄弟子块归并（2026-09-15） -------------------------------------------
|
||
|
||
|
||
def test_sibling_subblocks_of_one_section_collapse_to_the_highest_scoring_one() -> None:
|
||
"""同一节的多个子块是"同一答案的不同细节"，不是并列候选：只留最高分那条。"""
|
||
hits = [
|
||
hit("POL-AST-009-12", 0.7359),
|
||
hit("POL-AST-009-07", 0.7346),
|
||
hit("POL-AST-009-19", 0.7340),
|
||
hit("POL-AST-009-51", 0.7340),
|
||
]
|
||
|
||
merged = KnowledgeSearchService._merge_sibling_subblocks(hits)
|
||
|
||
assert [item.doc_id for item in merged] == ["POL-AST-009-12"]
|
||
|
||
|
||
def test_merge_keeps_one_block_per_section() -> None:
|
||
"""**不同**父块各自的最高分子块都要留下：它们是真正不同的候选。"""
|
||
hits = [
|
||
hit("PROD-007-04", 0.86),
|
||
hit("PROD-007-05", 0.85),
|
||
hit("HNW-005-02", 0.80),
|
||
hit("HNW-005-01", 0.79),
|
||
]
|
||
|
||
merged = KnowledgeSearchService._merge_sibling_subblocks(hits)
|
||
|
||
assert [item.doc_id for item in merged] == ["PROD-007-04", "HNW-005-02"]
|
||
|
||
|
||
def test_merge_leaves_plain_blocks_alone() -> None:
|
||
"""FAQ / 政策 / 公司信息这类块本身就是细粒度答案：它们之间打平是真的多个候选，
|
||
**不能**合并（否则"存在并列"这个信号会被抹掉，客服会硬答一个巧合高分）。"""
|
||
hits = [
|
||
hit("FAQ-0016", 0.85),
|
||
hit("FAQ-0015", 0.84),
|
||
hit("POL-SPM-010", 0.83),
|
||
hit("POL-AST-009", 0.82),
|
||
]
|
||
|
||
assert KnowledgeSearchService._merge_sibling_subblocks(hits) == hits
|
||
|
||
|
||
def test_merge_preserves_score_order_and_the_parent_block() -> None:
|
||
"""归并不改顺序；父块（整节）不受影响，仍会按保底名额回到候选里。"""
|
||
hits = [
|
||
hit("PROD-007-04", 0.86),
|
||
hit("PROD-007", 0.77), # 父块：整节，`_parent_of` 为 None
|
||
hit("PROD-007-05", 0.75),
|
||
hit("FAQ-0015", 0.60),
|
||
]
|
||
|
||
merged = KnowledgeSearchService._merge_sibling_subblocks(hits)
|
||
|
||
assert [item.doc_id for item in merged] == ["PROD-007-04", "PROD-007", "FAQ-0015"]
|
||
assert [item.score for item in merged] == [0.86, 0.77, 0.60]
|
||
|
||
|
||
def test_merge_is_idempotent() -> None:
|
||
"""重复调用不得继续删东西（幂等，便于以后在别处复用）。"""
|
||
hits = [hit("PROD-007-04", 0.86), hit("PROD-007-05", 0.75), hit("FAQ-0015", 0.60)]
|
||
|
||
once = KnowledgeSearchService._merge_sibling_subblocks(hits)
|
||
twice = KnowledgeSearchService._merge_sibling_subblocks(once)
|
||
|
||
assert twice == once
|
||
|
||
|
||
def test_merge_of_empty_list_is_empty() -> None:
|
||
assert KnowledgeSearchService._merge_sibling_subblocks([]) == []
|