feat(knowledge): 知识检索增加产品名的字面兜底召回
问题：客户问「季季盈90天的起投金额是多少」会被引导到人工客服，而知识库里明明有答案。 实测根因不是阈值拍错了，而是专有名词在 embedding 空间里不占优势——该问句的向量 top1 只有 0.6291，够不到 0.75 硬门槛，只能靠与次优的差值勉强通过；而同一次查询用 title like "%季季盈%" 是唯一命中 PROD-007。既然客户已经说出了产品名，就不该再赌相似度。 做法（三条边界都是实测逼出来的，不是设想）： 1. 只对产品集合做字面匹配。客户问「季季盈90天的起投金额是多少」与通用 FAQ 标题 「基金起投金额是多少?」有 7 个字连续重合；把 FAQ 纳入字面匹配会让它和真正的产品块 一起拿到满分、差距归零，反而又退化成"转人工"。 2. 字面命中只在向量结果不够确定时采用。客户问「基金赎回几天到账」时向量已给出正确答案 （FAQ-0016 得 0.8060），但手册章节标题「5.2 基金赎回流程」与问句也有 4 个字连续重合， 无条件采纳会把"操作步骤"顶掉客户真正问的"到账时间"。 3. 重叠门槛取 6 字而不是 4 字："基金赎回"这类业务动作词正好 4 字，会骗过 4 字门槛； 产品名（"南方季季盈90天"）更长，6 字能同时保住产品名、挡住动作词。 未改动任何转人工判定阈值；VECTOR_CONFIDENT_SCORE 与 Agent 的 HIGH_SCORE 由单测锁定一致， 避免两处各自漂移出"谁都答不出来"的死角。 验证：季季盈类问法由"转人工"变为正确答出，基金赎回问法仍答 FAQ-0016； ruff / mypy(113 文件) / 453 unit+contract / 29 integration 全绿。
This commit is contained in:
@@ -25,6 +25,30 @@ OUTPUT_FIELDS = (
|
||||
"tags", "doc_no", "version", "source_file", "visibility",
|
||||
)
|
||||
|
||||
# 关键字精确召回：客户问到产品名这类**专有名词**时，字面匹配比相似度更确定。
|
||||
#
|
||||
# 为什么需要：专有名词在 embedding 空间里不占优势。实测 160 条知识，客户问
|
||||
# 「季季盈90天的起投金额是多少」向量 top1 = PROD-007 仅 0.6291（够不到 0.75 的硬门槛，
|
||||
# 只能靠与次优的差值勉强通过）；而同一次查询用 `title like "%季季盈%"` 是**唯一命中**
|
||||
# PROD-007。既然客户已经明确说出了产品名，就不该再让相似度去赌。
|
||||
KEYWORD_MATCH_SCORE = 1.0 # 字面命中的确定分，压过任何相似度分
|
||||
# 与标题的最长公共子串至少要这么长，才算"客户确切提到了它"。
|
||||
# 为什么不是 4：实测客户问「基金赎回几天到账」，与手册章节标题「5.2 基金赎回流程」
|
||||
# 正好有 4 个字连续重合——但"基金赎回"是业务动作词，不是专有名词。产品名
|
||||
# （"南方季季盈90天"）通常比业务动作词长，取 6 字能同时保住产品名、挡住动作词。
|
||||
MIN_KEYWORD_OVERLAP = 6
|
||||
KEYWORD_SCAN_LIMIT = 500 # 一次最多扫描多少条标题；知识库到上千块后应改为倒排索引
|
||||
|
||||
# 字面匹配只在向量结果**不够确定**时介入。
|
||||
#
|
||||
# 这条门禁是实测逼出来的：客户问「基金赎回几天到账」，向量已给出正确答案
|
||||
# （FAQ-0016「基金赎回到账需要多长时间?」得 0.8060），但产品手册里的章节标题
|
||||
# 「五、申购赎回操作指南」与问句也有 6 个字连续重合，无条件字面匹配会把它顶到第一，
|
||||
# 用操作步骤替换掉客户真正问的到账时间。所以字面匹配是**兜底**，不是优先。
|
||||
#
|
||||
# 门槛必须与客服 Agent 的高置信门槛一致，tests/unit 有断言锁定两者相等。
|
||||
VECTOR_CONFIDENT_SCORE = 0.75
|
||||
|
||||
|
||||
class VectorSearcher(Protocol):
|
||||
"""只依赖用到的两个方法，便于测试替身注入。"""
|
||||
@@ -138,6 +162,12 @@ class KnowledgeSearchService:
|
||||
continue
|
||||
collected.extend(self._parse(raw, collection))
|
||||
|
||||
# 第二路召回：客户确切说出的产品名按字面取回。只在向量结果不够确定时介入，
|
||||
# 否则会把向量已经答对的题顶掉（见 VECTOR_CONFIDENT_SCORE 的说明）。
|
||||
best_vector_score = max((hit.score for hit in collected), default=0.0)
|
||||
if best_vector_score < VECTOR_CONFIDENT_SCORE:
|
||||
collected.extend(self._product_keyword_hits(client, targets, text, expression))
|
||||
|
||||
if not collected and failures == len(targets) and targets:
|
||||
# 三个集合全查失败：是链路故障，不是"知识库里没有"
|
||||
return KnowledgeSearchOutcome(
|
||||
@@ -160,6 +190,103 @@ class KnowledgeSearchService:
|
||||
searched_collections=targets,
|
||||
)
|
||||
|
||||
# ---- 关键字精确召回（字面匹配） ----
|
||||
|
||||
@staticmethod
|
||||
def _overlap_length(left: str, right: str) -> int:
|
||||
"""两段文本的最长公共子串长度。
|
||||
|
||||
用最长公共子串而不是分词：知识库标题是「南方科技有限公司 个人理财产品手册 ·
|
||||
2.1 南方季季盈90天」这种没有词边界的长串，任何分词器都得先养一份自定义词典，
|
||||
而词典会和手册一起过期。子串匹配不需要词典，手册改版也不会失效。
|
||||
"""
|
||||
if not left or not right:
|
||||
return 0
|
||||
previous = [0] * (len(right) + 1)
|
||||
best = 0
|
||||
for i in range(1, len(left) + 1):
|
||||
current = [0] * (len(right) + 1)
|
||||
for j in range(1, len(right) + 1):
|
||||
if left[i - 1] == right[j - 1]:
|
||||
current[j] = previous[j - 1] + 1
|
||||
if current[j] > best:
|
||||
best = current[j]
|
||||
previous = current
|
||||
return best
|
||||
|
||||
def _product_keyword_hits(
|
||||
self, client: Any, targets: Sequence[str], query: str, expression: str | None
|
||||
) -> list[KnowledgeHit]:
|
||||
"""客户确切说出某个产品名时，按字面把它取出来（兜底用）。
|
||||
|
||||
两处边界都是实测逼出来的：
|
||||
|
||||
1. **只对产品集合做**。客户问「季季盈90天的起投金额是多少」，与通用 FAQ 标题
|
||||
「基金起投金额是多少?」的最长公共子串有 7 个字；若把 FAQ 也纳入字面匹配，
|
||||
它会和真正的产品块一起拿到满分、差距归零，反而又退化成"转人工"。
|
||||
2. **命中不能无条件优先**。产品手册的标题里不只有产品名，还有章节名：客户问
|
||||
「基金赎回几天到账」时，「五、申购赎回操作指南」那一块与问句也有 6 个字连续
|
||||
重合。所以本方法产出什么是一回事，是否采用由调用方按"向量是否已经足够确定"
|
||||
决定（见 VECTOR_CONFIDENT_SCORE）。
|
||||
|
||||
失败一律返回空：关键字路径是**增益**，它坏了不能让整个检索变成故障。
|
||||
"""
|
||||
lookup = getattr(client, "query", None)
|
||||
if lookup is None or PRODUCT_COLLECTION not in targets:
|
||||
return [] # 客户端不支持标量查询（如测试替身），或本次没查产品集合
|
||||
try:
|
||||
rows = lookup(
|
||||
collection_name=PRODUCT_COLLECTION,
|
||||
filter=expression,
|
||||
output_fields=["doc_id", "title"],
|
||||
limit=KEYWORD_SCAN_LIMIT,
|
||||
)
|
||||
except Exception:
|
||||
return []
|
||||
|
||||
matched_ids = [
|
||||
str(row.get("doc_id") or "")
|
||||
for row in (rows if isinstance(rows, list) else [])
|
||||
if isinstance(row, dict)
|
||||
and self._overlap_length(query, str(row.get("title") or "")) >= MIN_KEYWORD_OVERLAP
|
||||
]
|
||||
matched_ids = [doc_id for doc_id in matched_ids if doc_id]
|
||||
if not matched_ids:
|
||||
return []
|
||||
|
||||
quoted = ", ".join(f'"{doc_id}"' for doc_id in matched_ids)
|
||||
try:
|
||||
details = lookup(
|
||||
collection_name=PRODUCT_COLLECTION,
|
||||
filter=f"doc_id in [{quoted}]",
|
||||
output_fields=list(OUTPUT_FIELDS),
|
||||
limit=len(matched_ids),
|
||||
)
|
||||
except Exception:
|
||||
return []
|
||||
|
||||
hits: list[KnowledgeHit] = []
|
||||
for row in details if isinstance(details, list) else []:
|
||||
if not isinstance(row, dict):
|
||||
continue
|
||||
content = str(row.get("content") or "")
|
||||
if not content:
|
||||
continue
|
||||
hits.append(KnowledgeHit(
|
||||
doc_id=str(row.get("doc_id") or ""),
|
||||
title=str(row.get("title") or ""),
|
||||
content=content,
|
||||
score=KEYWORD_MATCH_SCORE,
|
||||
source_file=str(row.get("source_file") or ""),
|
||||
visibility=str(row.get("visibility") or "public"),
|
||||
doc_no=str(row.get("doc_no") or ""),
|
||||
version=str(row.get("version") or ""),
|
||||
chapter=str(row.get("chapter") or ""),
|
||||
))
|
||||
# 一个产品名可能命中多个块（产品概览、费率表各一块）：全都保留，
|
||||
# 是不是"只有一个明确候选"交给上层的 gap 判定，这里不替它做选择。
|
||||
return hits
|
||||
|
||||
@staticmethod
|
||||
def _parse(raw: Any, collection: str) -> list[KnowledgeHit]:
|
||||
"""把 pymilvus 的 `[[{id, distance, entity}]]` 折叠成命中列表（纯函数，不抛异常）。"""
|
||||
|
||||
@@ -0,0 +1,169 @@
|
||||
"""知识检索的「字面兜底召回」单元测试。
|
||||
|
||||
用例全部来自实测，不是设想：
|
||||
|
||||
1. 客户问「季季盈90天的起投金额是多少」，纯向量 top1 只有 0.6291，够不到 0.75 门槛；
|
||||
而它与产品块标题的最长公共子串是 6 个字（就是产品名本身），字面匹配能唯一锁定。
|
||||
2. 客户问「基金赎回几天到账」，纯向量 top1 是 0.8060 的**正确答案**，但它与手册章节
|
||||
标题「5.2 基金赎回流程」也有 4 个字连续重合——所以字面匹配必须同时满足
|
||||
"重叠够长（6 字）"与"向量不够确定"两个条件，否则会把已经答对的题顶掉。
|
||||
"""
|
||||
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
|
||||
from app.service.knowledge_search_service import (
|
||||
MIN_KEYWORD_OVERLAP,
|
||||
PRODUCT_COLLECTION,
|
||||
VECTOR_CONFIDENT_SCORE,
|
||||
KnowledgeSearchService,
|
||||
)
|
||||
|
||||
PRODUCT_TITLE = "南方科技有限公司 个人理财产品手册 · 二、银行理财产品 · 2.1 南方季季盈90天"
|
||||
FLOW_TITLE = "南方科技有限公司 个人理财产品手册 · 五、申购赎回操作流程 · 5.2 基金赎回流程"
|
||||
|
||||
|
||||
async def _embed(text: str) -> list[float]:
|
||||
return [0.1, 0.2, 0.3]
|
||||
|
||||
|
||||
def _row(doc_id: str, title: str, score: float, content: str = "正文") -> dict[str, Any]:
|
||||
"""构造 pymilvus 的 `{distance, entity}` 行（`search()` 的返回形状）。"""
|
||||
return {
|
||||
"distance": score,
|
||||
"entity": {
|
||||
"doc_id": doc_id, "title": title, "content": content,
|
||||
"source_file": "x.md", "visibility": "public", "doc_no": "",
|
||||
"version": "", "chapter": "",
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
def _flat(doc_id: str, title: str, content: str = "正文") -> dict[str, Any]:
|
||||
"""构造 pymilvus 的扁平行（`query()` 的返回形状，字段直接挂在顶层）。
|
||||
|
||||
与 `search()` 的 `{distance, entity}` 不是同一种形状，所以这里刻意分成两个构造函数：
|
||||
本文件第一版把 `_row` 用在了标量查询上，于是业务代码解析出空 content 并跳过它，
|
||||
测试红了一次——是测试写错，不是业务代码有问题（业务代码"没有正文就不作答"是对的）。
|
||||
"""
|
||||
return {
|
||||
"doc_id": doc_id, "title": title, "content": content,
|
||||
"source_file": "x.md", "visibility": "public", "doc_no": "",
|
||||
"version": "", "chapter": "",
|
||||
}
|
||||
|
||||
|
||||
class FakeClient:
|
||||
"""向量检索与标量查询的替身；记录 query 调用次数以便断言"有没有走字面匹配"。"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vector_rows: list[dict[str, Any]],
|
||||
product_titles: list[tuple[str, str]],
|
||||
product_rows: dict[str, dict[str, Any]],
|
||||
) -> None:
|
||||
self._vector_rows = vector_rows
|
||||
self._product_titles = product_titles
|
||||
self._product_rows = product_rows
|
||||
self.query_calls = 0
|
||||
|
||||
def search(self, **kwargs: Any) -> Any:
|
||||
return [self._vector_rows]
|
||||
|
||||
def query(self, **kwargs: Any) -> Any:
|
||||
self.query_calls += 1
|
||||
if "content" not in kwargs.get("output_fields", []):
|
||||
return [{"doc_id": d, "title": t} for d, t in self._product_titles]
|
||||
pattern = str(kwargs.get("filter") or "")
|
||||
return [row for doc_id, row in self._product_rows.items() if doc_id in pattern]
|
||||
|
||||
|
||||
def _service(client: FakeClient) -> KnowledgeSearchService:
|
||||
return KnowledgeSearchService(client, _embed, collections=[PRODUCT_COLLECTION])
|
||||
|
||||
|
||||
def test_overlap_length_matches_measured_facts() -> None:
|
||||
"""锁定实测到的两段重叠长度，防止有人把阈值当成"拍脑袋的数"随手改掉。"""
|
||||
overlap = KnowledgeSearchService._overlap_length
|
||||
assert overlap("季季盈90天的起投金额是多少", PRODUCT_TITLE) == 6
|
||||
assert overlap("基金赎回几天到账", FLOW_TITLE) == 4
|
||||
assert overlap("客户想了解开户材料", PRODUCT_TITLE) == 0
|
||||
assert overlap("", PRODUCT_TITLE) == 0
|
||||
# 4 字的业务动作词必须被 6 字门槛挡在外面
|
||||
assert overlap("基金赎回几天到账", FLOW_TITLE) < MIN_KEYWORD_OVERLAP
|
||||
|
||||
|
||||
def test_gate_value_matches_agent_high_score() -> None:
|
||||
"""检索层的"向量够确定了"门槛与客服 Agent 的高置信门槛必须一致。
|
||||
|
||||
两处各自漂移的话，会出现"Agent 认为不够确定要转人工，检索层却认为够确定不给兜底"
|
||||
这种谁都答不出来的死角。
|
||||
"""
|
||||
from app.service.agent.implementations.customer_service import HIGH_SCORE
|
||||
|
||||
assert VECTOR_CONFIDENT_SCORE == HIGH_SCORE
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_literal_match_rescues_weak_vector_result() -> None:
|
||||
"""向量给不出高置信答案时，字面命中的产品块以确定分胜出（季季盈实测）。"""
|
||||
client = FakeClient(
|
||||
vector_rows=[_row("PROD-901", "某无关章节", 0.62)],
|
||||
product_titles=[("PROD-007", PRODUCT_TITLE)],
|
||||
product_rows={"PROD-007": _flat("PROD-007", PRODUCT_TITLE, "产品正文")},
|
||||
)
|
||||
outcome = await _service(client).search("季季盈90天的起投金额是多少")
|
||||
|
||||
assert outcome.hits[0].doc_id == "PROD-007"
|
||||
assert outcome.hits[0].score == 1.0
|
||||
assert client.query_calls == 2 # 先取标题表，再取命中块的正文
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_literal_match_stays_out_when_vector_is_confident() -> None:
|
||||
"""向量已给出高置信答案时，字面匹配不得介入（"基金赎回流程"实测反例）。"""
|
||||
client = FakeClient(
|
||||
vector_rows=[_row("FAQ-0016", "基金赎回到账需要多长时间?", 0.806)],
|
||||
product_titles=[("PROD-015", FLOW_TITLE)],
|
||||
product_rows={"PROD-015": _flat("PROD-015", FLOW_TITLE, "操作步骤")},
|
||||
)
|
||||
outcome = await _service(client).search("基金赎回几天到账")
|
||||
|
||||
assert outcome.hits[0].doc_id == "FAQ-0016"
|
||||
assert client.query_calls == 0 # 一次标量查询都不该发生
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_client_without_query_support_degrades_silently() -> None:
|
||||
"""客户端（如精简替身）不支持标量查询时，字面匹配静默跳过，不影响向量召回。"""
|
||||
|
||||
class NoQueryClient:
|
||||
def search(self, **kwargs: Any) -> Any:
|
||||
return [[_row("PROD-007", PRODUCT_TITLE, 0.62)]]
|
||||
|
||||
outcome = await KnowledgeSearchService(
|
||||
NoQueryClient(), _embed, collections=[PRODUCT_COLLECTION]
|
||||
).search("季季盈90天的起投金额是多少")
|
||||
|
||||
assert len(outcome.hits) == 1
|
||||
assert outcome.degraded is False
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_literal_lookup_failure_does_not_break_search() -> None:
|
||||
"""标量查询抛异常时字面匹配返回空，向量结果照常返回。"""
|
||||
|
||||
class BrokenQueryClient(FakeClient):
|
||||
def query(self, **kwargs: Any) -> Any:
|
||||
raise RuntimeError("milvus 标量查询挂了")
|
||||
|
||||
client = BrokenQueryClient(
|
||||
vector_rows=[_row("PROD-007", PRODUCT_TITLE, 0.62)],
|
||||
product_titles=[("PROD-007", PRODUCT_TITLE)],
|
||||
product_rows={},
|
||||
)
|
||||
outcome = await _service(client).search("季季盈90天的起投金额是多少")
|
||||
|
||||
assert [hit.doc_id for hit in outcome.hits] == ["PROD-007"]
|
||||
assert outcome.degraded is False
|
||||
Reference in New Issue
Block a user