更新 neo4j_indexing.py
检索索引构建冗余代码
This commit is contained in:
parent
eb07576567
commit
b919725117
@ -35,8 +35,6 @@ EMBED_MAX_WORKERS = SEARCH_CONFIG["embed_thread_num"] # embedding 批次并行
|
||||
NODE_FETCH_PAGE = SEARCH_CONFIG["node_fetch_page"] # 单次拉取待生成 embedding 节点的分页大小
|
||||
|
||||
# 全局索引名称（覆盖所有节点标签）
|
||||
GLOBAL_VECTOR_INDEX_NAME = SEARCH_CONFIG["global_entity_embedding"]
|
||||
GLOBAL_FULLTEXT_INDEX_NAME = SEARCH_CONFIG["global_entity_content_search"]
|
||||
DEFAULT_URI = NEO4J_CONFIG["uri"]
|
||||
DEFAULT_AUTH = (
|
||||
NEO4J_CONFIG["username"],
|
||||
@ -57,8 +55,6 @@ OLD_INDEX_NAMES: List[str] = [
|
||||
VECTOR_INDEX_NAME,
|
||||
FULLTEXT_INDEX_NAME,
|
||||
FULLTEXT_FIELD_INDEX_NAME,
|
||||
GLOBAL_VECTOR_INDEX_NAME,
|
||||
GLOBAL_FULLTEXT_INDEX_NAME,
|
||||
]
|
||||
|
||||
|
||||
@ -192,390 +188,6 @@ def drop_index_without_constraint(driver):
|
||||
driver.execute_query(f"drop index {record['name']} if exists")
|
||||
|
||||
|
||||
# --------- 创建全局混合索引（覆盖所有节点标签） ---------
|
||||
def create_global_indexes(driver, force_refresh=False):
|
||||
"""
|
||||
创建覆盖所有节点标签的全局向量索引和全文索引。
|
||||
与按标签创建的索引互补，支持跨标签统一检索。
|
||||
|
||||
Args:
|
||||
driver: Neo4j driver
|
||||
force_refresh: True=先删除旧索引再重建，False=IF NOT EXISTS 跳过已有索引
|
||||
"""
|
||||
if force_refresh:
|
||||
for index_name in (GLOBAL_VECTOR_INDEX_NAME, GLOBAL_FULLTEXT_INDEX_NAME):
|
||||
try:
|
||||
driver.execute_query(f"DROP INDEX {index_name} IF EXISTS")
|
||||
logger.info(f"强制刷新：已删除全局索引 {index_name}")
|
||||
except Exception as e:
|
||||
logger.warning(f"删除全局索引 {index_name} 时出错（可能不存在）: {e}")
|
||||
|
||||
# 全局向量索引（作用于所有节点的 embedding 属性）
|
||||
try:
|
||||
driver.execute_query(f"""
|
||||
CREATE VECTOR INDEX {GLOBAL_VECTOR_INDEX_NAME} IF NOT EXISTS
|
||||
FOR (n) ON n.embedding
|
||||
OPTIONS {{indexConfig: {{
|
||||
`vector.dimensions`: {vector_dim},
|
||||
`vector.similarity_function`: 'cosine'
|
||||
}}}}
|
||||
""")
|
||||
logger.info(f"全局向量索引 '{GLOBAL_VECTOR_INDEX_NAME}' 已就绪")
|
||||
except Exception as e:
|
||||
logger.error(f"创建全局向量索引失败: {e}")
|
||||
raise
|
||||
|
||||
# 全局全文索引（作用于所有节点的 名称 和 内容 属性）
|
||||
try:
|
||||
driver.execute_query(f"""
|
||||
CREATE FULLTEXT INDEX {GLOBAL_FULLTEXT_INDEX_NAME} IF NOT EXISTS
|
||||
FOR (n) ON EACH [n.名称, n.内容]
|
||||
""")
|
||||
logger.info(f"全局全文索引 '{GLOBAL_FULLTEXT_INDEX_NAME}' 已就绪")
|
||||
except Exception as e:
|
||||
logger.error(f"创建全局全文索引失败: {e}")
|
||||
raise
|
||||
|
||||
|
||||
# --------- 创建向量索引 ---------
|
||||
def vector_indexing(driver, label, force_refresh=False):
|
||||
"""
|
||||
创建向量索引（如果节点没有 embedding，使用"名称"属性生成 embedding）
|
||||
|
||||
说明：
|
||||
- 待生成 embedding 的节点通过分页循环全部处理，不再受单次 LIMIT 限制，
|
||||
避免节点数超过一万时被静默丢弃。
|
||||
- embedding 计算采用并行批次以缩短耗时。
|
||||
|
||||
Args:
|
||||
driver: Neo4j driver
|
||||
label: 节点标签
|
||||
force_refresh: 是否强制刷新（如果索引已存在，force_refresh=False 时会跳过，True 时会先删除再创建）
|
||||
"""
|
||||
|
||||
# === 分页处理所有缺失 embedding 的节点 ===
|
||||
# 每轮取一页未生成 embedding 的节点并写回，直到没有剩余节点为止。
|
||||
# 这样即使待处理节点远超一万也能全部覆盖。
|
||||
total_generated = 0
|
||||
while True:
|
||||
query_no_embedding = f"""
|
||||
MATCH (n:{label})
|
||||
WHERE n.embedding IS NULL
|
||||
AND n.名称 IS NOT NULL
|
||||
RETURN elementId(n) AS id, n.名称 AS 名称
|
||||
LIMIT {NODE_FETCH_PAGE}
|
||||
"""
|
||||
result_no_embedding = driver.execute_query(query_no_embedding)
|
||||
nodes_to_process = [
|
||||
(record["id"], record["名称"]) for record in result_no_embedding.records
|
||||
]
|
||||
|
||||
if not nodes_to_process:
|
||||
break
|
||||
|
||||
logger.info(f"标签 {label} 本轮发现 {len(nodes_to_process)} 个节点需要生成 embedding")
|
||||
|
||||
ids, names = zip(*nodes_to_process)
|
||||
ids = list(ids)
|
||||
names = list(names)
|
||||
|
||||
# 并行计算 embedding
|
||||
logger.info(f"开始为 {len(names)} 个节点计算 embedding（并行批次）...")
|
||||
try:
|
||||
all_embeddings = _compute_embeddings_parallel(names)
|
||||
except Exception as e:
|
||||
logger.error(f"批量嵌入向量计算失败: {e}")
|
||||
raise
|
||||
|
||||
# 转换为 numpy 数组并做 L2 归一化
|
||||
embeddings = np.array(all_embeddings)
|
||||
norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
|
||||
norms = np.where(norms == 0, 1, norms) # 避免除零
|
||||
embeddings = embeddings / norms
|
||||
|
||||
# 保存 embedding 到节点
|
||||
logger.info(f"写入 {len(ids)} 个节点的 embedding...")
|
||||
upsert_vectors(
|
||||
driver,
|
||||
ids=ids,
|
||||
embedding_property="embedding",
|
||||
embeddings=embeddings,
|
||||
)
|
||||
total_generated += len(ids)
|
||||
logger.info(f"本轮成功为 {len(ids)} 个节点生成并保存 embedding")
|
||||
|
||||
# 不足一页说明已处理完，提前结束
|
||||
if len(nodes_to_process) < NODE_FETCH_PAGE:
|
||||
break
|
||||
|
||||
if total_generated:
|
||||
logger.info(f"标签 {label} 共生成 {total_generated} 个节点的 embedding")
|
||||
|
||||
# 检查是否有节点存在 embedding（包括刚生成的）
|
||||
check_query = f"MATCH (n:{label}) WHERE n.embedding IS NOT NULL RETURN count(n) AS count"
|
||||
result = driver.execute_query(check_query)
|
||||
node_count = result.records[0]["count"] if result.records else 0
|
||||
|
||||
if node_count == 0:
|
||||
logger.warning(f"标签 {label} 没有节点包含 embedding 属性，跳过向量索引创建")
|
||||
return
|
||||
|
||||
# 创建向量索引
|
||||
index_name = f"{label.lower()}_vector"
|
||||
|
||||
# 如果强制刷新且索引已存在，先删除它
|
||||
if force_refresh:
|
||||
try:
|
||||
driver.execute_query(f"DROP INDEX {index_name} IF EXISTS")
|
||||
logger.info(f"强制刷新模式：已删除旧索引 {index_name}")
|
||||
except Exception as e:
|
||||
logger.warning(f"删除索引 {index_name} 时出错（可能不存在）: {e}")
|
||||
|
||||
try:
|
||||
create_vector_index(
|
||||
driver,
|
||||
name=index_name, # 索引的唯一名称
|
||||
label=label, # 要索引的节点标签
|
||||
embedding_property="embedding", # 包含嵌入向量值的节点属性键
|
||||
dimensions=vector_dim, # 向量嵌入维度，1024与使用的 bge-m3 嵌入模型一致
|
||||
similarity_fn="cosine", # 向量相似度函数，可选 "euclidean" 或 "cosine"
|
||||
)
|
||||
logger.info(f"成功为标签 {label} 创建向量索引（共 {node_count} 个节点）")
|
||||
except Exception as e:
|
||||
# 如果是增量更新模式且索引已存在，这是正常的，只记录警告
|
||||
if not force_refresh and "already exists" in str(e).lower():
|
||||
logger.info(f"标签 {label} 的向量索引已存在，跳过创建")
|
||||
else:
|
||||
logger.error(f"为标签 {label} 创建向量索引失败: {e}")
|
||||
raise
|
||||
|
||||
|
||||
# --------- 创建全文索引 ---------
|
||||
def fulltext_indexing(driver, label, force_refresh=False):
|
||||
"""
|
||||
创建全文索引，并添加节点属性（自动获取所有属性）
|
||||
|
||||
全文索引：词 ---》 id
|
||||
后续检索，分词后，根据词反向找到节点
|
||||
|
||||
Args:
|
||||
driver: Neo4j driver
|
||||
label: 节点标签
|
||||
force_refresh: 是否强制刷新所有节点的全文索引（True=清空后重新生成，False=只处理缺失的节点）
|
||||
"""
|
||||
# 排除内部属性
|
||||
excluded_props = {"embedding", "fulltext", "elementId", "id"}
|
||||
|
||||
# 根据 force_refresh 决定查询条件
|
||||
if force_refresh:
|
||||
# 强制刷新：处理所有节点（包括已有 fulltext 的节点）
|
||||
query = f"""MATCH (n:{label})
|
||||
WITH n, elementId(n) AS id, properties(n) AS props
|
||||
RETURN id, props"""
|
||||
else:
|
||||
# 增量更新：只处理没有 fulltext 的节点
|
||||
query = f"""MATCH (n:{label})
|
||||
WHERE n.fulltext IS NULL
|
||||
WITH n, elementId(n) AS id, properties(n) AS props
|
||||
RETURN id, props"""
|
||||
|
||||
# 执行查询
|
||||
records = driver.execute_query(query).records
|
||||
|
||||
if not records:
|
||||
if force_refresh:
|
||||
logger.info(f"{label} 没有节点需要处理")
|
||||
else:
|
||||
logger.info(f"{label} 所有节点皆存在全文索引属性")
|
||||
return
|
||||
|
||||
# 构建文本描述（使用所有属性）
|
||||
record_list = []
|
||||
for r in records:
|
||||
node_id = r["id"]
|
||||
props = r["props"]
|
||||
# 过滤掉内部属性
|
||||
filtered_props = {k: v for k, v in props.items() if k not in excluded_props}
|
||||
# 使用所有属性构建文本
|
||||
text = build_text_from_all_properties(filtered_props)
|
||||
if text:
|
||||
record_list.append({"id": node_id, "text": text})
|
||||
|
||||
if not record_list:
|
||||
logger.info(f"{label} 没有可索引的文本内容")
|
||||
return
|
||||
|
||||
# 创建全文索引（在写入数据之前创建索引）
|
||||
try:
|
||||
create_fulltext_index(
|
||||
driver,
|
||||
name=f"{label.lower()}_fulltext", # 索引的唯一名称
|
||||
label=label, # 要创建索引的节点标签
|
||||
node_properties=["fulltext"], # 要创建全文索引的节点属性列表
|
||||
)
|
||||
except Exception as e:
|
||||
logger.warning(f"创建全文索引可能已存在（继续处理）: {e}")
|
||||
|
||||
# 文本分词，作为全文索引属性
|
||||
logger.info(f"计算 {label} ({len(record_list)}) 的全文索引")
|
||||
pattern = re.compile(r"[a-zA-Z0-9\u4e00-\u9fa5]+") # 匹配英文字母、数字和中文字符
|
||||
fulltext_tuple_list = []
|
||||
for id_, text in [(r["id"], r["text"]) for r in record_list]:
|
||||
# 使用 jieba 分词
|
||||
words = jieba.lcut(text)
|
||||
# 过滤并连接
|
||||
filtered_words = [
|
||||
word.strip()
|
||||
for word in words
|
||||
if word.strip() and pattern.fullmatch(word.strip())
|
||||
]
|
||||
fulltext_value = " ".join(filtered_words)
|
||||
if fulltext_value:
|
||||
fulltext_tuple_list.append((id_, fulltext_value))
|
||||
|
||||
if not fulltext_tuple_list:
|
||||
logger.warning(f"{label} 分词后没有有效内容")
|
||||
return
|
||||
|
||||
ids, fulltexts = zip(*fulltext_tuple_list)
|
||||
ids = list(ids)
|
||||
fulltexts = list(fulltexts)
|
||||
|
||||
# 按 elementId 添加全文索引属性
|
||||
logger.info(f"写入 {label} ({len(fulltexts)}) 的全文索引")
|
||||
insert_batch_size = 1000
|
||||
for i in range(0, len(fulltext_tuple_list), insert_batch_size):
|
||||
batch_rows = [
|
||||
{"id": id_, "fulltext": ft}
|
||||
for id_, ft in zip(
|
||||
ids[i: i + insert_batch_size],
|
||||
fulltexts[i: i + insert_batch_size],
|
||||
)
|
||||
]
|
||||
|
||||
# UNWIND：将列表数据展开为多行记录
|
||||
driver.execute_query(
|
||||
"UNWIND $rows AS row " # 将传入的 rows 列表展开，每项作为一行数据，命名为 row
|
||||
"MATCH (n) "
|
||||
"WHERE elementId(n) = row.id "
|
||||
"SET n.fulltext = row.fulltext ",
|
||||
{"rows": batch_rows},
|
||||
)
|
||||
|
||||
|
||||
def create_all_indexes(neo4j_uri=None, neo4j_username=None, neo4j_password=None, force_refresh=False, driver=None):
|
||||
"""
|
||||
创建所有索引的主函数（可被外部调用）
|
||||
|
||||
注意：Neo4j 连接配置优先从 .env 文件中读取（NEO4J_URI, NEO4J_USERNAME, NEO4J_PASSWORD）
|
||||
如果 .env 文件中没有配置，则使用参数中的值（如果提供）或默认值
|
||||
|
||||
Args:
|
||||
neo4j_uri: Neo4j URI（可选，优先级低于 .env，提供 driver 时忽略）
|
||||
neo4j_username: Neo4j 用户名（可选，优先级低于 .env，提供 driver 时忽略）
|
||||
neo4j_password: Neo4j 密码（可选，优先级低于 .env，提供 driver 时忽略）
|
||||
force_refresh: 是否强制刷新所有索引（True=清空重建，False=增量更新）
|
||||
driver: 可选的 Neo4j driver 实例，提供则复用外部 driver
|
||||
|
||||
Returns:
|
||||
dict: 包含执行结果的字典
|
||||
"""
|
||||
use_external_driver = driver is not None
|
||||
|
||||
if not use_external_driver:
|
||||
# 优先从 .env 文件读取配置，如果 .env 中没有则使用参数或默认值
|
||||
neo4j_uri = os.getenv("NEO4J_URI") or neo4j_uri or "bolt://192.168.0.46:57687"
|
||||
neo4j_username = os.getenv("NEO4J_USERNAME") or neo4j_username or "neo4j"
|
||||
neo4j_password = os.getenv("NEO4J_PASSWORD") or neo4j_password or "zdht123@"
|
||||
logger.info(f"连接到 Neo4j: {neo4j_uri}")
|
||||
else:
|
||||
logger.info("使用外部提供的 Neo4j driver")
|
||||
|
||||
# 定义内部函数来执行索引创建逻辑（避免代码重复）
|
||||
def _execute_indexing(driver):
|
||||
# 1、根据 force_refresh 决定是否清空索引和约束
|
||||
if force_refresh:
|
||||
logger.info("强制刷新模式：清空所有约束...")
|
||||
drop_constraint(driver)
|
||||
logger.info("强制刷新模式：清空所有索引...")
|
||||
drop_index_without_constraint(driver)
|
||||
logger.info("强制刷新模式：清空所有节点的 fulltext 属性...")
|
||||
driver.execute_query("MATCH (n) WHERE n.fulltext IS NOT NULL REMOVE n.fulltext")
|
||||
else:
|
||||
logger.info("增量更新模式：保留现有索引和约束")
|
||||
|
||||
# 2、自动获取所有节点标签并创建向量索引
|
||||
logger.info("=" * 50)
|
||||
logger.info("开始创建向量索引...")
|
||||
logger.info("=" * 50)
|
||||
|
||||
all_labels = get_all_labels(driver)
|
||||
logger.info(f"发现 {len(all_labels)} 个节点标签: {all_labels}")
|
||||
|
||||
vector_index_errors = []
|
||||
for label in all_labels:
|
||||
logger.info(f"处理标签: {label}")
|
||||
try:
|
||||
vector_indexing(driver, label, force_refresh=force_refresh)
|
||||
except Exception as e:
|
||||
error_msg = f"为标签 {label} 创建向量索引时出错: {e}"
|
||||
logger.error(error_msg)
|
||||
vector_index_errors.append(error_msg)
|
||||
continue
|
||||
|
||||
# 3、创建全文索引（自动获取所有属性）
|
||||
logger.info("=" * 50)
|
||||
logger.info("开始创建全文索引...")
|
||||
logger.info("=" * 50)
|
||||
|
||||
fulltext_index_errors = []
|
||||
for label in all_labels:
|
||||
logger.info(f"处理标签: {label}")
|
||||
try:
|
||||
fulltext_indexing(driver, label, force_refresh=force_refresh)
|
||||
except Exception as e:
|
||||
error_msg = f"为标签 {label} 创建全文索引时出错: {e}"
|
||||
logger.error(error_msg)
|
||||
fulltext_index_errors.append(error_msg)
|
||||
continue
|
||||
|
||||
logger.info("=" * 50)
|
||||
logger.info("所有索引创建完成！")
|
||||
logger.info("=" * 50)
|
||||
|
||||
has_errors = len(vector_index_errors) > 0 or len(fulltext_index_errors) > 0
|
||||
mode_text = "强制刷新" if force_refresh else "增量更新"
|
||||
return {
|
||||
"success": not has_errors,
|
||||
"message": f"{mode_text}模式：索引创建完成" if not has_errors
|
||||
else f"{mode_text}模式：索引创建完成，但部分标签出现错误",
|
||||
"mode": "force_refresh" if force_refresh else "incremental",
|
||||
"labels": all_labels,
|
||||
"vector_index_errors": vector_index_errors,
|
||||
"fulltext_index_errors": fulltext_index_errors,
|
||||
}
|
||||
|
||||
try:
|
||||
if use_external_driver:
|
||||
# 使用外部 driver，直接执行，不关闭 driver
|
||||
return _execute_indexing(driver)
|
||||
else:
|
||||
# 使用内部 driver，使用上下文管理器自动关闭
|
||||
with GraphDatabase.driver(neo4j_uri, auth=(neo4j_username, neo4j_password)) as driver:
|
||||
return _execute_indexing(driver)
|
||||
except Exception as e:
|
||||
error_msg = f"创建索引过程中发生错误: {e}"
|
||||
logger.error(error_msg, exc_info=True)
|
||||
return {
|
||||
"success": False,
|
||||
"message": error_msg,
|
||||
"mode": "error",
|
||||
"labels": [],
|
||||
"vector_index_errors": [],
|
||||
"fulltext_index_errors": [],
|
||||
}
|
||||
|
||||
|
||||
def print_neo4j_version(driver) -> None:
|
||||
"""打印当前 Neo4j 实例的版本和版本类型。"""
|
||||
with driver.session() as session:
|
||||
@ -795,13 +407,12 @@ build_hyrid_indexes = build_hybrid_indexes
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
# 可以通过命令行参数或环境变量控制是否强制刷新
|
||||
import sys
|
||||
|
||||
force_refresh = "--force" in sys.argv or os.getenv("FORCE_REFRESH", "false").lower() == "true"
|
||||
result = build_hybrid_indexes(
|
||||
drop_old=True
|
||||
)
|
||||
|
||||
result = create_all_indexes(force_refresh=force_refresh)
|
||||
if result["success"]:
|
||||
logger.info(f"索引创建成功完成（模式: {result['mode']}）")
|
||||
logger.info("混合索引创建成功")
|
||||
else:
|
||||
logger.error(f"索引创建完成但有错误: {result['message']}")
|
||||
logger.error(result["message"])
|
||||
Loading…
x
Reference in New Issue
Block a user