[Question]:PowerMem query speed is extremely slow, 6-7 seconds

[Hasen-dc]

Description

PowerMem查询速度极慢，6-7秒，而且库里面只有不到10条数据。大模型（Qwen/Qwen3-4B-Instruct-2507-FP8）和嵌入模型（Qwen/Qwen3-Embedding-0.6B）都是使用的参数极小的，向量库使用oceanbase（1024维）。三个模块分别使用脚本测试，时延都是正常的，毫秒级，但组合在一起，就延时这么多，不知道哪里的问题。

使用 https://www.powermem.ai/docs/examples/scenario_10_sparse_vector 的例子得到的结果如下：

✓ Configuration updated from YAML successfully
✓ Current Mode: UserMemory (Profile Enabled)
✓ Memory initialized successfully
Adding test memories...

Search query: 'AI algorithms'
Search time: 4.074261426925659s
Found 5 results:

1. Score: 0.0164
   Content: Machine learning is a branch of artificial intelli...

2. Score: 0.0081
   Content: Deep learning uses multi-layer neural networks for...

3. Score: 0.0079
   Content: Natural language processing is an interdisciplinar...

4. Score: 0.0078
   Content: Knowledge graphs are graph-structured data used to...

5. Score: 0.0077
   Content: Vector search is an important technology for infor...

---

Description

PowerMem query speed is extremely slow, 6-7 seconds, and there are less than 10 pieces of data in the database. The large model (Qwen/Qwen3-4B-Instruct-2507-FP8) and the embedded model (Qwen/Qwen3-Embedding-0.6B) both use extremely small parameters, and the vector library uses oceanbase (1024 dimensions). The three modules were tested separately using scripts, and the delays were all normal, at the millisecond level, but when combined together, the delay was so much that I don’t know what the problem is.

The results obtained using the example of https://www.powermem.ai/docs/examples/scenario_10_sparse_vector are as follows:

✓ Configuration updated from YAML successfully
✓ Current Mode: UserMemory (Profile Enabled)
✓ Memory initialized successfully
Adding test memories...

Search query: 'AI algorithms'
Search time: 4.074261426925659s
Found 5 results:

1. Score: 0.0164
   Content: Machine learning is a branch of artificial intelligence...

2. Score: 0.0081
   Content: Deep learning uses multi-layer neural networks for...

3. Score: 0.0079
   Content: Natural language processing is an interdisciplinar...

4. Score: 0.0078
   Content: Knowledge graphs are graph-structured data used to...

5. Score: 0.0077
   Content: Vector search is an important technology for infor...

[Evenss]

感谢您的反馈。我们并未能复现出该情况，能提供具体的env文件/json配置文件（隐藏掉关键的ak信息）和数据库的具体版本吗？

---

Thank you for your feedback. We have not been able to reproduce this situation. Can you provide the specific env file/json configuration file (which hides the key ak information) and the specific version of the database?

[Hasen-dc]

感谢您的反馈。我们并未能复现出该情况，能提供具体的env文件/json配置文件（隐藏掉关键的ak信息）和数据库的具体版本吗？

Thank you for your feedback. We have not been able to reproduce this situation. Can you provide the specific env file/json configuration file (which hides the key ak information) and the specific version of the database?

配置信息如下：

`config = {
# 启用用户画像模式 (UserMemory)
'enable_user_profile': True,

# ========== LLM 配置 ==========
'llm': {
    'provider': 'openai',
    'config': {
        'api_key': 'test_key',
        'model': 'Qwen/Qwen3-4B-Instruct-2507-FP8',
        'openai_base_url': 'http://x.x.x.x:yy/v1',
        'temperature': 0.2  # 保留原代码中的温度设置
    }
},

# ========== Embedding 配置 ==========
'embedder': {
    'provider': 'openai',
    'config': {
        'api_key': 'test_key',
        'model': 'Qwen/Qwen3-Embedding-0.6B',
        'openai_base_url': 'http://x2.x2.x2.x2:yy/v1',
        'embedding_dims': 1024
    }
},

# ========== Database 配置 (OceanBase) ==========
'vector_store': {
    'provider': 'oceanbase',
    'config': {
        'host': 'x3.x3.x3.x3',
        'port': 2881,
        'user': 'test',
        'password': 'password',
        'db_name': 'db_name',
        'collection_name': 'collection_name',
        'embedding_model_dims': 1024,
        'include_sparse': True
    }
}

}`

数据库的信息：
+------------------------------+
| @@Version |
+------------------------------+
| 5.7.25-OceanBase_CE-v4.3.5.5 |
+------------------------------+

CPU 服务器：8核/16G/60G系统盘/240G数据盘

docker 方式部署

下面的测试程序，开了infer=False，可以都加入数据库，注释掉这个配置，会出现只能将第一条记忆加入数据库，另外两条返回为空。

`memory = Memory(config=config)
user_id = "user123"
memory_array = ["User works as a software engineer", "User likes Python programming", "User prefers email support"]

for content in memory_array:
t1 = time.time()
ret = memory.add(
messages=content,
user_id=user_id,
# infer=False # Do not use intelligent inference, store directly
)
t2 = time.time()
print(f"Added {content} in {t2-t1}s, ret = {ret}")
time.sleep(3)
`
log信息如下：

Added User works as a software engineer in 1.9583380222320557s, ret = {'results': [{'id': 683969592258723840, 'memory': 'User is a software engineer', 'event': 'ADD'}]} Added User likes Python programming in 1.1771209239959717s, ret = {'results': []} Added User prefers email support in 1.1025419235229492s, ret = {'results': []}