mirror of
https://github.com/xinnan-tech/xiaozhi-esp32-server.git
synced 2026-07-27 17:43:55 +08:00
Merge branch 'xinnan-tech:main' into custom_tts_api
This commit is contained in:
hsoftxl
@@ -231,7 +231,9 @@ class ConnectionHandler:
|
||||
# 创建新事件循环（避免与主循环冲突）
|
||||
loop = asyncio.new_event_loop()
|
||||
asyncio.set_event_loop(loop)
|
||||
loop.run_until_complete(self.memory.save_memory(self.dialogue.dialogue))
|
||||
loop.run_until_complete(
|
||||
self.memory.save_memory(self.dialogue.dialogue)
|
||||
)
|
||||
except Exception as e:
|
||||
self.logger.bind(tag=TAG).error(f"保存记忆失败: {e}")
|
||||
finally:
|
||||
@@ -441,10 +443,10 @@ class ConnectionHandler:
|
||||
def _initialize_memory(self):
|
||||
"""初始化记忆模块"""
|
||||
self.memory.init_memory(
|
||||
self.device_id,
|
||||
self.llm,
|
||||
self.config["summaryMemory"],
|
||||
not self.read_config_from_api,
|
||||
role_id=self.device_id,
|
||||
llm=self.llm,
|
||||
summary_memory=self.config.get("summaryMemory", None),
|
||||
save_to_file=not self.read_config_from_api,
|
||||
)
|
||||
|
||||
def _initialize_intent(self):
|
||||
|
||||
@@ -215,7 +215,7 @@ class IntentProvider(IntentProviderBase):
|
||||
|
||||
# 记录识别到的function call
|
||||
logger.bind(tag=TAG).info(
|
||||
f"识别到function call: {function_name}, 参数: {function_args}"
|
||||
f"llm 识别到意图: {function_name}, 参数: {function_args}"
|
||||
)
|
||||
|
||||
# 添加到缓存
|
||||
|
||||
@@ -21,27 +21,67 @@ class LLMProvider(LLMProviderBase):
|
||||
api_key="ollama" # Ollama doesn't need an API key but OpenAI client requires one
|
||||
)
|
||||
|
||||
# 检查是否是qwen3模型
|
||||
self.is_qwen3 = self.model_name and self.model_name.lower().startswith("qwen3")
|
||||
|
||||
def response(self, session_id, dialogue):
|
||||
try:
|
||||
# 如果是qwen3模型，在用户最后一条消息中添加/no_think指令
|
||||
if self.is_qwen3:
|
||||
# 复制对话列表，避免修改原始对话
|
||||
dialogue_copy = dialogue.copy()
|
||||
|
||||
# 找到最后一条用户消息
|
||||
for i in range(len(dialogue_copy) - 1, -1, -1):
|
||||
if dialogue_copy[i]["role"] == "user":
|
||||
# 在用户消息前添加/no_think指令
|
||||
dialogue_copy[i]["content"] = "/no_think " + dialogue_copy[i]["content"]
|
||||
logger.bind(tag=TAG).debug(f"为qwen3模型添加/no_think指令")
|
||||
break
|
||||
|
||||
# 使用修改后的对话
|
||||
dialogue = dialogue_copy
|
||||
|
||||
responses = self.client.chat.completions.create(
|
||||
model=self.model_name,
|
||||
messages=dialogue,
|
||||
stream=True
|
||||
)
|
||||
is_active=True
|
||||
is_active = True
|
||||
# 用于处理跨chunk的标签
|
||||
buffer = ""
|
||||
|
||||
for chunk in responses:
|
||||
try:
|
||||
delta = chunk.choices[0].delta if getattr(chunk, 'choices', None) else None
|
||||
content = delta.content if hasattr(delta, 'content') else ''
|
||||
|
||||
if content:
|
||||
if '<think>' in content:
|
||||
# 将内容添加到缓冲区
|
||||
buffer += content
|
||||
|
||||
# 处理缓冲区中的标签
|
||||
while '<think>' in buffer and '</think>' in buffer:
|
||||
# 找到完整的<think></think>标签并移除
|
||||
pre = buffer.split('<think>', 1)[0]
|
||||
post = buffer.split('</think>', 1)[1]
|
||||
buffer = pre + post
|
||||
|
||||
# 处理只有开始标签的情况
|
||||
if '<think>' in buffer:
|
||||
is_active = False
|
||||
content = content.split('<think>')[0]
|
||||
if '</think>' in content:
|
||||
buffer = buffer.split('<think>', 1)[0]
|
||||
|
||||
# 处理只有结束标签的情况
|
||||
if '</think>' in buffer:
|
||||
is_active = True
|
||||
content = content.split('</think>')[-1]
|
||||
if is_active:
|
||||
yield content
|
||||
buffer = buffer.split('</think>', 1)[1]
|
||||
|
||||
# 如果当前处于活动状态且缓冲区有内容，则输出
|
||||
if is_active and buffer:
|
||||
yield buffer
|
||||
buffer = "" # 清空缓冲区
|
||||
|
||||
except Exception as e:
|
||||
logger.bind(tag=TAG).error(f"Error processing chunk: {e}")
|
||||
|
||||
@@ -51,6 +91,22 @@ class LLMProvider(LLMProviderBase):
|
||||
|
||||
def response_with_functions(self, session_id, dialogue, functions=None):
|
||||
try:
|
||||
# 如果是qwen3模型，在用户最后一条消息中添加/no_think指令
|
||||
if self.is_qwen3:
|
||||
# 复制对话列表，避免修改原始对话
|
||||
dialogue_copy = dialogue.copy()
|
||||
|
||||
# 找到最后一条用户消息
|
||||
for i in range(len(dialogue_copy) - 1, -1, -1):
|
||||
if dialogue_copy[i]["role"] == "user":
|
||||
# 在用户消息前添加/no_think指令
|
||||
dialogue_copy[i]["content"] = "/no_think " + dialogue_copy[i]["content"]
|
||||
logger.bind(tag=TAG).debug(f"为qwen3模型添加/no_think指令")
|
||||
break
|
||||
|
||||
# 使用修改后的对话
|
||||
dialogue = dialogue_copy
|
||||
|
||||
stream = self.client.chat.completions.create(
|
||||
model=self.model_name,
|
||||
messages=dialogue,
|
||||
@@ -58,8 +114,49 @@ class LLMProvider(LLMProviderBase):
|
||||
tools=functions,
|
||||
)
|
||||
|
||||
is_active = True
|
||||
buffer = ""
|
||||
|
||||
for chunk in stream:
|
||||
yield chunk.choices[0].delta.content, chunk.choices[0].delta.tool_calls
|
||||
try:
|
||||
delta = chunk.choices[0].delta if getattr(chunk, 'choices', None) else None
|
||||
content = delta.content if hasattr(delta, 'content') else None
|
||||
tool_calls = delta.tool_calls if hasattr(delta, 'tool_calls') else None
|
||||
|
||||
# 如果是工具调用，直接传递
|
||||
if tool_calls:
|
||||
yield None, tool_calls
|
||||
continue
|
||||
|
||||
# 处理文本内容
|
||||
if content:
|
||||
# 将内容添加到缓冲区
|
||||
buffer += content
|
||||
|
||||
# 处理缓冲区中的标签
|
||||
while '<think>' in buffer and '</think>' in buffer:
|
||||
# 找到完整的<think></think>标签并移除
|
||||
pre = buffer.split('<think>', 1)[0]
|
||||
post = buffer.split('</think>', 1)[1]
|
||||
buffer = pre + post
|
||||
|
||||
# 处理只有开始标签的情况
|
||||
if '<think>' in buffer:
|
||||
is_active = False
|
||||
buffer = buffer.split('<think>', 1)[0]
|
||||
|
||||
# 处理只有结束标签的情况
|
||||
if '</think>' in buffer:
|
||||
is_active = True
|
||||
buffer = buffer.split('</think>', 1)[1]
|
||||
|
||||
# 如果当前处于活动状态且缓冲区有内容，则输出
|
||||
if is_active and buffer:
|
||||
yield buffer, None
|
||||
buffer = "" # 清空缓冲区
|
||||
except Exception as e:
|
||||
logger.bind(tag=TAG).error(f"Error processing function chunk: {e}")
|
||||
continue
|
||||
|
||||
except Exception as e:
|
||||
logger.bind(tag=TAG).error(f"Error in Ollama function call: {e}")
|
||||
|
||||
@@ -21,6 +21,6 @@ class MemoryProviderBase(ABC):
|
||||
"""Query memories for specific role based on similarity"""
|
||||
return "please implement query method"
|
||||
|
||||
def init_memory(self, role_id, llm, summary_memory=None):
|
||||
def init_memory(self, role_id, llm, **kwargs):
|
||||
self.role_id = role_id
|
||||
self.llm = llm
|
||||
|
||||
@@ -112,8 +112,10 @@ class MemoryProvider(MemoryProviderBase):
|
||||
self.memory_path = get_project_dir() + "data/.memory.yaml"
|
||||
self.load_memory(summary_memory)
|
||||
|
||||
def init_memory(self, role_id, llm, summary_memory=None, save_to_file=True):
|
||||
super().init_memory(role_id, llm)
|
||||
def init_memory(
|
||||
self, role_id, llm, summary_memory=None, save_to_file=True, **kwargs
|
||||
):
|
||||
super().init_memory(role_id, llm, **kwargs)
|
||||
self.save_to_file = save_to_file
|
||||
self.load_memory(summary_memory)
|
||||
|
||||
|
||||
@@ -319,7 +319,7 @@ def initialize_modules(
|
||||
modules["memory"] = memory.create_instance(
|
||||
memory_type,
|
||||
config["Memory"][select_memory_module],
|
||||
config['summaryMemory'],
|
||||
config.get('summaryMemory', None),
|
||||
)
|
||||
logger.bind(tag=TAG).info(f"初始化组件: memory成功 {select_memory_module}")
|
||||
|
||||
|
||||
@@ -76,6 +76,7 @@ services:
|
||||
expose:
|
||||
- 6379
|
||||
container_name: xiaozhi-esp32-server-redis
|
||||
restart: always
|
||||
healthcheck:
|
||||
test: ["CMD", "redis-cli", "ping"]
|
||||
interval: 10s
|
||||
|
||||
@@ -22,11 +22,12 @@ mem0ai==0.1.62
|
||||
bs4==0.0.2
|
||||
modelscope==1.23.2
|
||||
sherpa_onnx==1.11.0
|
||||
mcp==1.7.1
|
||||
mcp==1.8.1
|
||||
cnlunar==0.2.0
|
||||
PySocks==1.7.1
|
||||
dashscope==1.23.1
|
||||
baidu-aip==4.16.13
|
||||
chardet==5.2.0
|
||||
aioconsole==0.8.1
|
||||
markitdown==0.1.1
|
||||
markitdown==0.1.1
|
||||
mcp-proxy==0.6.0
|
||||
|
||||
Reference in New Issue
Block a user