Hass/xiaozhi-esp32-server
1
0
Fork 0
mirror of https://github.com/xinnan-tech/xiaozhi-esp32-server.git synced 2026-07-22 07:03:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #2822 from wayyoungboy/add-powermem

Browse Source 
fix: powermem sqlite支持用户画像以及修改配置模式
This commit is contained in:
Sakura-RanChen
2026-01-09 14:46:31 +08:00
committed by GitHub
parent ad93f43145 06a2e6fbfa
commit f66966d72f
4 changed files with 239 additions and 181 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/powermem-integration.md
+129 -105
View File
@@ -46,36 +46,58 @@ selected_module:
Memory:
powermem:
type: powermem
# 是否启用用户画像功能（需要OceanBase）
enable_user_profile: false
# 数据库提供者: oceanbase(推荐,最佳性能), seekdb, postgres, sqlite(轻量备选)
database_provider: sqlite # 资源充足时建议使用 oceanbase 或 seekdb
# LLM提供者: qwen(默认), openai, 等
llm_provider: qwen
# 嵌入模型提供者: qwen(默认), openai, 等
embedding_provider: qwen
# LLM配置
llm_api_key: 你的LLM API密钥
llm_model: qwen-plus
# 嵌入模型配置
embedding_api_key: 你的嵌入模型API密钥
embedding_model: text-embedding-v3
# 是否启用用户画像功能
# 用户画像支持: oceanbase、seekdb、sqlite (powermem 0.3.0+)
enable_user_profile: true
# ========== LLM 配置 ==========
llm:
provider: openai # 可选: qwen, openai, zhipu 等
config:
api_key: 你的LLM API密钥
model: qwen-plus
# openai_base_url: https://api.openai.com/v1 # 可选，自定义服务地址
# ========== Embedding 配置 ==========
embedder:
provider: openai # 可选: qwen, openai 等
config:
api_key: 你的嵌入模型API密钥
model: text-embedding-v4
openai_base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
# ========== Database 配置 ==========
vector_store:
provider: sqlite # 可选: oceanbase(推荐), seekdb(推荐), postgres, sqlite(轻量)
config: {} # SQLite 无需额外配置
```
### 配置参数详解
| 参数 | 说明 | 默认值 | 可选值 |
|------|------|--------|--------|
| `enable_user_profile` | 启用用户画像模式 | `false` | `true`(需OceanBase), `false` |
| `database_provider` | 存储后端类型 | `sqlite` | `oceanbase`(推荐), `seekdb`, `postgres`, `sqlite`(轻量) |
| `llm_provider` | LLM 提供商 | `qwen` | `qwen`, `zhipu`(免费), `openai`, 等 |
| `embedding_provider` | 嵌入模型提供商 | `qwen` | `qwen`, `zhipu`, `openai`, 等 |
| `llm_api_key` | LLM API 密钥 | - | - |
| `llm_model` | LLM 模型名称 | - | 根据提供商选择 |
| `llm_base_url` | LLM API 地址（可选） | - | - |
| `embedding_api_key` | 嵌入模型 API 密钥 | - | - |
| `embedding_model` | 嵌入模型名称 | - | 根据提供商选择 |
| `embedding_base_url` | 嵌入模型 API 地址（可选） | - | - |
#### LLM 配置
| 参数 | 说明 | 可选值 |
|------|------|--------|
| `llm.provider` | LLM 提供商 | `qwen`, `openai`, `zhipu` 等 |
| `llm.config.api_key` | API 密钥 | - |
| `llm.config.model` | 模型名称 | 根据提供商选择 |
| `llm.config.openai_base_url` | 自定义服务地址（可选） | - |
#### Embedding 配置
| 参数 | 说明 | 可选值 |
|------|------|--------|
| `embedder.provider` | 嵌入模型提供商 | `qwen`, `openai` 等 |
| `embedder.config.api_key` | API 密钥 | - |
| `embedder.config.model` | 模型名称 | 根据提供商选择 |
| `embedder.config.openai_base_url` | 自定义服务地址（可选） | - |
#### Database 配置
| 参数 | 说明 | 可选值 |
|------|------|--------|
| `vector_store.provider` | 存储后端类型 | `oceanbase`(推荐), `seekdb`(推荐), `postgres`, `sqlite`(轻量) |
| `vector_store.config` | 数据库连接配置 | 根据 provider 设置 |
### 记忆模式说明
@@ -84,7 +106,9 @@ PowerMem 支持两种记忆模式：
| 模式 | 配置 | 功能 | 存储要求 |
|------|------|------|----------|
| **普通记忆** | `enable_user_profile: false` | 对话记忆存储与检索 | 支持所有数据库 |
| **用户画像** | `enable_user_profile: true` | 记忆 + 自动提取用户画像 | 仅支持 OceanBase |
| **用户画像** | `enable_user_profile: true` | 记忆 + 自动提取用户画像 | oceanbase、seekdb、sqlite |
> 📌 **版本说明**：PowerMem 0.3.0+ 版本，用户画像功能支持 OceanBase、SeekDB、SQLite 三种存储后端。
### 使用通义千问（推荐）
@@ -96,13 +120,21 @@ PowerMem 支持两种记忆模式：
Memory:
powermem:
type: powermem
database_provider: sqlite
llm_provider: qwen
embedding_provider: qwen
llm_api_key: sk-xxxxxxxxxxxxxxxx
llm_model: qwen-plus
embedding_api_key: sk-xxxxxxxxxxxxxxxx
embedding_model: text-embedding-v3
enable_user_profile: true
llm:
provider: qwen
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: qwen-plus
embedder:
provider: openai
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: text-embedding-v4
openai_base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
vector_store:
provider: sqlite
config: {}
```
### 使用智谱免费 LLM（完全免费方案）
@@ -117,15 +149,22 @@ Memory:
Memory:
powermem:
type: powermem
database_provider: sqlite
llm_provider: zhipu
embedding_provider: zhipu
llm_api_key: xxxxxxxxxxxxxxxx.xxxxxxxxxxxxxxxx
llm_model: glm-4-flash
llm_base_url: https://open.bigmodel.cn/api/paas/v4/
embedding_api_key: xxxxxxxxxxxxxxxx.xxxxxxxxxxxxxxxx
embedding_model: embedding-3
embedding_base_url: https://open.bigmodel.cn/api/paas/v4/
enable_user_profile: true
llm:
provider: openai # 使用 openai 兼容模式
config:
api_key: xxxxxxxxxxxxxxxx.xxxxxxxxxxxxxxxx
model: glm-4-flash
openai_base_url: https://open.bigmodel.cn/api/paas/v4/
embedder:
provider: openai
config:
api_key: xxxxxxxxxxxxxxxx.xxxxxxxxxxxxxxxx
model: embedding-3
openai_base_url: https://open.bigmodel.cn/api/paas/v4/
vector_store:
provider: sqlite
config: {}
```
### 使用 OpenAI
@@ -134,15 +173,22 @@ Memory:
Memory:
powermem:
type: powermem
database_provider: sqlite
llm_provider: openai
embedding_provider: openai
llm_api_key: sk-xxxxxxxxxxxxxxxx
llm_model: gpt-4o-mini
llm_base_url: https://api.openai.com/v1
embedding_api_key: sk-xxxxxxxxxxxxxxxx
embedding_model: text-embedding-3-small
embedding_base_url: https://api.openai.com/v1
enable_user_profile: true
llm:
provider: openai
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: gpt-4o-mini
openai_base_url: https://api.openai.com/v1
embedder:
provider: openai
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: text-embedding-3-small
openai_base_url: https://api.openai.com/v1
vector_store:
provider: sqlite
config: {}
```
### 使用 OceanBase（最佳性能方案）
@@ -158,14 +204,18 @@ OceanBase 是 PowerMem 的最佳搭档，可实现最大性能释放：
Memory:
powermem:
type: powermem
database_provider: oceanbase
llm_provider: qwen
embedding_provider: qwen
llm_api_key: sk-xxxxxxxxxxxxxxxx
llm_model: qwen-plus
embedding_api_key: sk-xxxxxxxxxxxxxxxx
embedding_model: text-embedding-v3
# OceanBase 数据库连接配置
enable_user_profile: true
llm:
provider: qwen
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: qwen-plus
embedder:
provider: openai
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: text-embedding-v4
openai_base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
vector_store:
provider: oceanbase
config:
@@ -173,35 +223,9 @@ Memory:
port: 2881
user: root@test
password: your_password
database: powermem
```
### 高级配置
如果需要更精细的控制，可以使用完整的配置结构：
```yaml
Memory:
powermem:
type: powermem
# 向量存储配置
vector_store:
provider: sqlite
config:
path: ./data/powermem.db
# LLM 配置
llm:
provider: qwen
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: qwen-plus
# 嵌入模型配置
embedder:
provider: qwen
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: text-embedding-v3
db_name: powermem
collection_name: memories # 默认值
embedding_model_dims: 1536 # 嵌入向量维度，必需参数
```
## 设备记忆隔离
@@ -216,6 +240,8 @@ PowerMem 会自动使用设备 ID（`device_id`）作为 `user_id` 进行记忆
PowerMem 提供 `UserMemory` 类，可自动从对话中提取用户画像信息。
> 📌 **版本说明**：PowerMem 0.3.0+ 版本，用户画像功能支持 OceanBase、SeekDB、SQLite 三种存储后端。
### 启用用户画像
在配置中设置 `enable_user_profile: true` 即可启用：
@@ -225,22 +251,20 @@ Memory:
powermem:
type: powermem
enable_user_profile: true # 启用用户画像
database_provider: oceanbase # 必须使用 OceanBase
llm_provider: qwen
embedding_provider: qwen
llm_api_key: sk-xxxxxxxxxxxxxxxx
llm_model: qwen-plus
embedding_api_key: sk-xxxxxxxxxxxxxxxx
embedding_model: text-embedding-v3
# OceanBase 数据库连接配置
vector_store:
provider: oceanbase
llm:
provider: qwen
config:
host: 127.0.0.1
port: 2881
user: root@test
password: your_password
database: powermem
api_key: sk-xxxxxxxxxxxxxxxx
model: qwen-plus
embedder:
provider: openai
config:
api_key: sk-xxxxxxxxxxxxxxxx
model: text-embedding-v4
openai_base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
vector_store:
provider: sqlite # 用户画像支持: oceanbase、seekdb、sqlite
config: {}
```
### 用户画像能力
@@ -258,7 +282,7 @@ Memory:
1. **用户画像**：用户的基本信息、兴趣爱好等
2. **相关记忆**：与当前对话相关的历史记忆
> ⚠️ **注意**：`UserMemory` 功能需要 OceanBase 作为存储后端，其他数据库暂不支持。
> ✅ **版本说明**：PowerMem 0.3.0+ 版本，用户画像功能支持 OceanBase、SeekDB、SQLite 三种存储后端。
## 与其他记忆组件的对比
main/xiaozhi-server/config.yaml
+38 -18
View File
@@ -284,26 +284,46 @@ Memory:
# 官网: https://www.powermem.ai/
# 使用示例: https://github.com/oceanbase/powermem/tree/main/examples
type: powermem
# 是否启用用户画像功能（需要OceanBase作为存储后端）
# - false: 使用普通记忆模式(AsyncMemory)，支持所有数据库
# - true: 使用用户画像模式(UserMemory)，自动提取用户信息，仅支持OceanBase
enable_user_profile: false
# 数据库提供者: oceanbase(推荐,最佳性能), seekdb(推荐,AI应用存储一体), postgres, sqlite(轻量备选)
# 资源充足时建议使用 oceanbase 或 seekdb
# 注意：用户画像功能(enable_user_profile: true)仅支持oceanbase
database_provider: sqlite
# LLM提供者: qwen(默认), zhipu(免费glm-4-flash), openai, 等
llm_provider: qwen
# 嵌入模型提供者: qwen(默认), zhipu, openai, 等
embedding_provider: qwen
# LLM配置 (使用通义千问为例)
# 是否启用用户画像功能
# - false: 使用普通记忆模式(AsyncMemory)
# - true: 使用用户画像模式(UserMemory)，自动提取用户信息
# 用户画像功能支持: oceanbase、seekdb、sqlite (powermem 0.3.0+)
enable_user_profile: true
# ========== LLM 配置 ==========
# 通义千问: https://bailian.console.aliyun.com/?apiKey=1#/api-key
# 智谱AI(免费): https://bigmodel.cn/usercenter/proj-mgmt/apikeys
llm_api_key: 你的LLM API密钥
llm_model: qwen-plus
# 嵌入模型配置 (使用通义千问为例)
embedding_api_key: 你的嵌入模型API密钥
embedding_model: text-embedding-v3
llm:
provider: openai # 可选: qwen, openai, zhipu 等
config:
api_key: 你的LLM API密钥
model: qwen-plus
# openai_base_url: https://api.openai.com/v1 # 可选，自定义服务地址
# ========== Embedding 配置 ==========
embedder:
provider: openai # 可选: qwen, openai 等
config:
api_key: 你的嵌入模型API密钥
model: text-embedding-v4
openai_base_url: https://dashscope.aliyuncs.com/compatible-mode/v1
# ========== Database 配置 ==========
# oceanbase(推荐,最佳性能), seekdb(推荐,AI应用存储一体), postgres, sqlite(轻量备选)
# 用户画像功能支持: oceanbase、seekdb、sqlite
vector_store:
provider: sqlite
config: {} # SQLite 无需额外配置
# OceanBase 配置示例:
# provider: oceanbase
# config:
# host: 127.0.0.1
# port: 2881
# user: root@test
# password: your_password
# db_name: powermem
# collection_name: memories # 默认值
# embedding_model_dims: 1536 # 嵌入向量维度，必需参数
nomem:
# 不想使用记忆功能，可以使用nomem
type: nomem
main/xiaozhi-server/core/providers/memory/powermem/powermem.py
+71 -57
View File
@@ -10,6 +10,7 @@
@Author: wayyoungboy
"""
import asyncio
import traceback
from typing import Optional, Dict, Any
@@ -21,13 +22,13 @@ TAG = __name__
class MemoryProvider(MemoryProviderBase):
"""
PowerMem memory provider implementation.
PowerMem is an open-source agent memory component that provides
efficient memory management for AI agents.
Supports multiple storage backends (sqlite, oceanbase, postgres),
LLM providers (qwen, openai, etc.) and embedding providers.
Config options:
- enable_user_profile: bool - Enable UserMemory for user profiling (requires OceanBase)
- database_provider: str - Storage backend (sqlite, oceanbase, postgres)
@@ -40,7 +41,8 @@ class MemoryProvider(MemoryProviderBase):
self.use_powermem = False
self.memory_client = None
self.enable_user_profile = False
self.last_profile_content = "" # Cache for user profile from UserMemory
try:
# Check if user profile mode is enabled
self.enable_user_profile = config.get("enable_user_profile", False)
@@ -49,21 +51,13 @@ class MemoryProvider(MemoryProviderBase):
database_provider = config.get("database_provider", "sqlite")
llm_provider = config.get("llm_provider", "qwen")
embedding_provider = config.get("embedding_provider", "qwen")
# UserMemory requires OceanBase
if self.enable_user_profile and database_provider not in ["oceanbase"]:
logger.bind(tag=TAG).warning(
f"UserMemory requires OceanBase as storage backend, but got {database_provider}. "
"Falling back to AsyncMemory mode."
)
self.enable_user_profile = False
# Build powermem configuration dict
# PowerMem supports two config styles:
# 1. powermem style: database, llm, embedding
# 2. mem0 style: vector_store, llm, embedder
powermem_config = {}
# Configure vector store / database
if "vector_store" in config:
powermem_config["vector_store"] = config["vector_store"]
@@ -74,7 +68,7 @@ class MemoryProvider(MemoryProviderBase):
"provider": database_provider,
"config": {}
}
# Configure LLM
if "llm" in config:
powermem_config["llm"] = config["llm"]
@@ -84,13 +78,23 @@ class MemoryProvider(MemoryProviderBase):
llm_config["api_key"] = config["llm_api_key"]
if "llm_model" in config:
llm_config["model"] = config["llm_model"]
# Handle base_url based on provider type
# - qwen provider uses dashscope_base_url
# - openai provider uses openai_base_url
if "llm_base_url" in config:
llm_config["base_url"] = config["llm_base_url"]
if llm_provider == "qwen":
llm_config["dashscope_base_url"] = config["llm_base_url"]
else:
llm_config["openai_base_url"] = config["llm_base_url"]
if "openai_base_url" in config:
llm_config["openai_base_url"] = config["openai_base_url"]
if "dashscope_base_url" in config:
llm_config["dashscope_base_url"] = config["dashscope_base_url"]
powermem_config["llm"] = {
"provider": llm_provider,
"config": llm_config
}
# Configure embedder
if "embedder" in config:
powermem_config["embedder"] = config["embedder"]
@@ -100,13 +104,25 @@ class MemoryProvider(MemoryProviderBase):
embedder_config["api_key"] = config["embedding_api_key"]
if "embedding_model" in config:
embedder_config["model"] = config["embedding_model"]
# Handle base_url based on provider type
# - qwen provider uses dashscope_base_url
# - openai provider uses openai_base_url
# Priority: embedding_xxx_base_url > embedding_base_url > xxx_base_url
if "embedding_base_url" in config:
embedder_config["base_url"] = config["embedding_base_url"]
if embedding_provider == "qwen":
embedder_config["dashscope_base_url"] = config["embedding_base_url"]
else:
embedder_config["openai_base_url"] = config["embedding_base_url"]
# Embedding-specific base_url (higher priority)
if "embedding_openai_base_url" in config:
embedder_config["openai_base_url"] = config["embedding_openai_base_url"]
if "embedding_dashscope_base_url" in config:
embedder_config["dashscope_base_url"] = config["embedding_dashscope_base_url"]
powermem_config["embedder"] = {
"provider": embedding_provider,
"config": embedder_config
}
# Initialize memory client based on mode
if self.enable_user_profile:
from powermem import UserMemory
@@ -116,7 +132,7 @@ class MemoryProvider(MemoryProviderBase):
from powermem import AsyncMemory
self.memory_client = AsyncMemory(config=powermem_config)
memory_mode = "AsyncMemory (普通记忆模式)"
self.use_powermem = True
logger.bind(tag=TAG).info(
@@ -140,6 +156,7 @@ class MemoryProvider(MemoryProviderBase):
Args:
msgs: List of message objects with 'role' and 'content' attributes
session_id: Session identifier (optional, for compatibility)
Returns:
@@ -148,7 +165,7 @@ class MemoryProvider(MemoryProviderBase):
if not self.use_powermem or self.memory_client is None:
logger.bind(tag=TAG).warning("PowerMem is not available, skipping save_memory")
return None
if len(msgs) < 2:
logger.bind(tag=TAG).debug("Not enough messages to save (need at least 2)")
return None
@@ -160,16 +177,26 @@ class MemoryProvider(MemoryProviderBase):
for message in msgs
if message.role != "system"
]
# Add memory using PowerMem SDK
result = await self.memory_client.add(
result = self.memory_client.add(
messages=messages,
user_id=self.role_id
)
# Handle both sync and async returns
if asyncio.iscoroutine(result):
result = await result
logger.bind(tag=TAG).debug(f"Save memory result: {result}")
# Cache user profile if UserMemory mode and profile was extracted
if self.enable_user_profile and result:
if result.get('profile_extracted'):
self.last_profile_content = result.get('profile_content', '')
logger.bind(tag=TAG).debug(f"User profile extracted: {self.last_profile_content}")
return result
except Exception as e:
logger.bind(tag=TAG).error(f"Error saving memory: {str(e)}")
logger.bind(tag=TAG).debug(f"Detailed error: {traceback.format_exc()}")
@@ -178,24 +205,24 @@ class MemoryProvider(MemoryProviderBase):
async def query_memory(self, query: str) -> str:
"""
Query memories from PowerMem based on similarity search.
Args:
query: The search query string
Returns:
Formatted string of relevant memories or empty string if none found
"""
if not self.use_powermem or self.memory_client is None:
logger.bind(tag=TAG).warning("PowerMem is not available, skipping query_memory")
return ""
try:
if not getattr(self, "role_id", None):
logger.bind(tag=TAG).debug("No role_id set, returning empty memory")
return ""
result_parts = []
# If user profile mode is enabled, include user profile in results
if self.enable_user_profile:
profile = await self.get_user_profile()
@@ -208,7 +235,7 @@ class MemoryProvider(MemoryProviderBase):
user_id=self.role_id,
limit=30
)
if results and "results" in results:
# Format each memory entry with its update time
memories = []
@@ -219,7 +246,7 @@ class MemoryProvider(MemoryProviderBase):
timestamp = str(entry["updated_at"])
elif "created_at" in entry and entry["created_at"]:
timestamp = str(entry["created_at"])
if timestamp:
try:
# Parse and reformat the timestamp (remove milliseconds if present)
@@ -232,7 +259,7 @@ class MemoryProvider(MemoryProviderBase):
formatted_time = timestamp
else:
formatted_time = ""
memory = entry.get("memory", "") or entry.get("content", "")
if memory:
if formatted_time:
@@ -252,7 +279,7 @@ class MemoryProvider(MemoryProviderBase):
final_result = "\n\n".join(result_parts)
logger.bind(tag=TAG).debug(f"Query results: {final_result}")
return final_result
except Exception as e:
logger.bind(tag=TAG).error(f"Error querying memory: {str(e)}")
logger.bind(tag=TAG).debug(f"Detailed error: {traceback.format_exc()}")
@@ -262,35 +289,22 @@ class MemoryProvider(MemoryProviderBase):
"""
Get user profile from PowerMem (only available in UserMemory mode).
In PowerMem 0.3.0+, user profile is automatically extracted during add()
and cached in last_profile_content.
Returns:
Formatted user profile string or empty string if not available
"""
if not self.use_powermem or self.memory_client is None:
return ""
if not self.enable_user_profile:
logger.bind(tag=TAG).debug("User profile mode is not enabled")
return ""
try:
if not getattr(self, "role_id", None):
return ""
# Get user profile using UserMemory SDK
profile = await self.memory_client.get_profile(user_id=self.role_id)
if not profile:
return ""
# Format profile as readable string
profile_parts = []
for key, value in profile.items():
if value:
profile_parts.append(f"- {key}: {value}")
return "\n".join(profile_parts)
except Exception as e:
logger.bind(tag=TAG).error(f"Error getting user profile: {str(e)}")
logger.bind(tag=TAG).debug(f"Detailed error: {traceback.format_exc()}")
return ""
# Return cached profile content from last add() operation
if self.last_profile_content:
return self.last_profile_content
return ""
main/xiaozhi-server/requirements.txt
+1 -1
View File
@@ -23,7 +23,7 @@ loguru==0.7.3
requests==2.32.5
cozepy==0.20.0
mem0ai==1.0.0
powermem>=0.2.1
powermem>=0.3.0
bs4==0.0.2
modelscope==1.32.0
sherpa_onnx==1.12.17