修复知识库搜索粒度:parent section是表格时优先返回child sections

- 添加_build_parent_child_map方法构建section层级关系
- 当parent section有heading命中但body是表格时,优先返回child sections
- 最终结果按分数排序确保最高分排第一
- 修复'安装说明'查询返回插件表格而非下载链接的问题
This commit is contained in:
youpaishidifu
2026-08-06 17:17:20 +08:00
parent 1291a402de
commit f84ea56311
+98
View File
@@ -384,6 +384,44 @@ class KnowledgeBase:
pairs.append((t, tokens[i + 1]))
return pairs
def _build_parent_child_map(self, doc_idx_list: List[int]) -> dict:
"""构建 section 的 parent-child 关系映射。
根据 heading 层级(# / ## / ### / ####)推断父子关系。
返回 {child_idx: parent_idx} 映射。
"""
import re
parent_map = {}
current_parent = {1: None, 2: None, 3: None, 4: None}
for idx in doc_idx_list:
if idx >= len(self.docs):
continue
doc = self.docs[idx]
heading = doc.get("heading", "")
# 提取 heading 层级
match = re.match(r'^(#{1,6})\s', heading)
if not match:
continue
level = len(match.group(1))
# 找到当前级别及以上的 parent
parent = None
for l in range(level - 1, 0, -1):
if current_parent.get(l) is not None:
parent = current_parent[l]
break
parent_map[idx] = parent
current_parent[level] = idx
# 重置更低级别的 parent
for l in range(level + 1, 5):
current_parent[l] = None
return parent_map
def search(self, query: str, top_k: int = 3, min_score: float = 0.1) -> List[Tuple[dict, float]]:
"""搜索知识库,返回 (doc, score) 列表。
@@ -721,6 +759,64 @@ class KnowledgeBase:
reranked.sort(key=lambda x: x[2], reverse=True)
doc_sections = reranked
# ⭐ Parent-child 精度提升:当 parent section 有 heading 命中但 body 是"空壳"(表格/短文本),
# 优先返回 child sections(更具体的子内容)
# 如 "## 2.安装说明:" body是前置插件表格,应优先返回 "### 2.1 下载" 子sections
if len(doc_sections) > 1:
# 构建当前文档的 parent-child 关系(从全部docs中)
all_doc_indices = list(range(len(self.docs)))
parent_map = self._build_parent_child_map(all_doc_indices)
# 检查是否有"空壳"parent section
for i, (idx, doc, score) in enumerate(doc_sections):
heading = doc.get("heading", "")
body = doc.get("body", "")
body_len = len(body.strip())
# 判断是否为"空壳"section:heading含查询词 + body是表格
h_tokens = set(t for t in jieba.cut(heading) if len(t) >= 2)
heading_has_query = bool(query_tokens & h_tokens)
is_table = body.strip().startswith('|') and '|' in body[1:]
if heading_has_query and is_table:
# 找到该section的children(从全部docs中)
children = [c_idx for c_idx, p_idx in parent_map.items() if p_idx == idx]
if children:
# 找到有实质性body的children
good_children = []
for child_idx in children:
child_doc = self.docs[child_idx]
child_body = child_doc.get("body", "")
child_body_len = len(child_body.strip())
child_heading = child_doc.get("heading", "")
# child有非空body且不是表格 → 候选
if child_body_len > 5 and not child_body.strip().startswith('|'):
good_children.append((child_idx, child_doc))
if good_children:
# 给parent减分(body是空壳表格)
doc_sections[i] = (idx, doc, score * 0.3)
# 将good children加入doc_sections(如果不在的话)
existing_idx_set = {s[0] for s in doc_sections}
for child_idx, child_doc in good_children:
if child_idx not in existing_idx_set:
# 计算child的加分:heading含查询词 → 额外加分
child_heading = child_doc.get("heading", "")
child_h_tokens = set(t for t in jieba.cut(child_heading) if len(t) >= 2)
child_heading_bonus = len(query_tokens & child_h_tokens) * 5.0
# child body含查询词 → 额外加分
child_body = child_doc.get("body", "")
child_body_bonus = 10.0 if any(t in child_body for t in query_tokens if len(t) >= 2) else 0.0
child_score = score * 0.8 + child_heading_bonus + child_body_bonus
doc_sections.append((child_idx, child_doc, child_score))
# 重新排序
doc_sections.sort(key=lambda x: x[2], reverse=True)
break
# ⭐ 表格优先策略(reranking后):条件优先返回与查询语义相关的表格section
# 防止reranking后表格被其他section超越(表格是答案但不含查询词)
# ⚠ 修复:不能盲目优先含"操作"的表格——"操作方式"表(交互模式)会覆盖"权限级别"表
@@ -908,6 +1004,8 @@ class KnowledgeBase:
break
results.append((doc, float(score)))
# ⭐ 最终排序:按分数降序,确保最高分的结果排第一
results.sort(key=lambda x: x[1], reverse=True)
return results[:top_k]
if results:
return results