Hass/xiaozhi-esp32-server
1
0
Fork 0
mirror of https://github.com/xinnan-tech/xiaozhi-esp32-server.git synced 2026-07-29 13:33:57 +08:00
Code Issues Packages Projects Releases Wiki Activity

update:智控台添加独立记忆模型配置

Browse Source
This commit is contained in:
hrz
2025-05-27 15:28:17 +08:00
parent 626692df29
commit be7ef08f40
11 changed files with 125 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files
main/xiaozhi-server/core/providers/llm/coze/coze.py
+1 -1
View File
@@ -25,7 +25,7 @@ class LLMProvider(LLMProviderBase):
self.session_conversation_map = {} # 存储session_id和conversation_id的映射
check_model_key("CozeLLM", self.personal_access_token)
def response(self, session_id, dialogue):
def response(self, session_id, dialogue, **kwargs):
coze_api_token = self.personal_access_token
coze_api_base = COZE_CN_BASE_URL
main/xiaozhi-server/core/providers/llm/dify/dify.py
+1 -1
View File
@@ -17,7 +17,7 @@ class LLMProvider(LLMProviderBase):
self.session_conversation_map = {} # 存储session_id和conversation_id的映射
check_model_key("DifyLLM", self.api_key)
def response(self, session_id, dialogue):
def response(self, session_id, dialogue, **kwargs):
try:
# 取最后一条用户消息
last_msg = next(m for m in reversed(dialogue) if m["role"] == "user")
main/xiaozhi-server/core/providers/llm/fastgpt/fastgpt.py
+1 -1
View File
@@ -16,7 +16,7 @@ class LLMProvider(LLMProviderBase):
self.variables = config.get("variables", {})
check_model_key("FastGPTLLM", self.api_key)
def response(self, session_id, dialogue):
def response(self, session_id, dialogue, **kwargs):
try:
# 取最后一条用户消息
last_msg = next(m for m in reversed(dialogue) if m["role"] == "user")
main/xiaozhi-server/core/providers/llm/gemini/gemini.py
+1 -1
View File
@@ -112,7 +112,7 @@ class LLMProvider(LLMProviderBase):
]
# Gemini文档提到，无需维护session-id，直接用dialogue拼接而成
def response(self, session_id, dialogue):
def response(self, session_id, dialogue, **kwargs):
yield from self._generate(dialogue, None)
def response_with_functions(self, session_id, dialogue, functions=None):
main/xiaozhi-server/core/providers/llm/homeassistant/homeassistant.py
+1 -1
View File
@@ -14,7 +14,7 @@ class LLMProvider(LLMProviderBase):
self.base_url = config.get("base_url", config.get("url")) # 默认使用 base_url
self.api_url = f"{self.base_url}/api/conversation/process" # 拼接完整的 API URL
def response(self, session_id, dialogue):
def response(self, session_id, dialogue, **kwargs):
try:
# home assistant语音助手自带意图，无需使用xiaozhi ai自带的，只需要把用户说的话传递给home assistant即可
main/xiaozhi-server/core/providers/llm/ollama/ollama.py
+38 -26
View File
@@ -18,13 +18,13 @@ class LLMProvider(LLMProviderBase):
self.client = OpenAI(
base_url=self.base_url,
api_key="ollama" # Ollama doesn't need an API key but OpenAI client requires one
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):
def response(self, session_id, dialogue, **kwargs):
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:
@@ -35,7 +35,9 @@ class LLMProvider(LLMProviderBase):
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"]
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
@@ -43,9 +45,7 @@ class LLMProvider(LLMProviderBase):
dialogue = dialogue_copy
responses = self.client.chat.completions.create(
model=self.model_name,
messages=dialogue,
stream=True
model=self.model_name, messages=dialogue, stream=True
)
is_active = True
# 用于处理跨chunk的标签
@@ -53,29 +53,33 @@ class LLMProvider(LLMProviderBase):
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 ''
delta = (
chunk.choices[0].delta
if getattr(chunk, "choices", None)
else None
)
content = delta.content if hasattr(delta, "content") else ""
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:
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]
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:
if "<think>" in buffer:
is_active = False
buffer = buffer.split('<think>', 1)[0]
buffer = buffer.split("<think>", 1)[0]
# 处理只有结束标签的情况
if '</think>' in buffer:
if "</think>" in buffer:
is_active = True
buffer = buffer.split('</think>', 1)[1]
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:
@@ -100,7 +104,9 @@ class LLMProvider(LLMProviderBase):
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"]
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
@@ -119,9 +125,15 @@ class LLMProvider(LLMProviderBase):
for chunk in stream:
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
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:
@@ -134,21 +146,21 @@ class LLMProvider(LLMProviderBase):
buffer += content
# 处理缓冲区中的标签
while '<think>' in buffer and '</think>' in buffer:
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]
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:
if "<think>" in buffer:
is_active = False
buffer = buffer.split('<think>', 1)[0]
buffer = buffer.split("<think>", 1)[0]
# 处理只有结束标签的情况
if '</think>' in buffer:
if "</think>" in buffer:
is_active = True
buffer = buffer.split('</think>', 1)[1]
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:
main/xiaozhi-server/core/providers/llm/xinference/xinference.py
+34 -21
View File
@@ -15,39 +15,45 @@ class LLMProvider(LLMProviderBase):
# 如果没有v1，增加v1
if not self.base_url.endswith("/v1"):
self.base_url = f"{self.base_url}/v1"
logger.bind(tag=TAG).info(f"Initializing Xinference LLM provider with model: {self.model_name}, base_url: {self.base_url}")
logger.bind(tag=TAG).info(
f"Initializing Xinference LLM provider with model: {self.model_name}, base_url: {self.base_url}"
)
try:
self.client = OpenAI(
base_url=self.base_url,
api_key="xinference" # Xinference has a similar setup to Ollama where it doesn't need an actual key
api_key="xinference", # Xinference has a similar setup to Ollama where it doesn't need an actual key
)
logger.bind(tag=TAG).info("Xinference client initialized successfully")
except Exception as e:
logger.bind(tag=TAG).error(f"Error initializing Xinference client: {e}")
raise
def response(self, session_id, dialogue):
def response(self, session_id, dialogue, **kwargs):
try:
logger.bind(tag=TAG).debug(f"Sending request to Xinference with model: {self.model_name}, dialogue length: {len(dialogue)}")
responses = self.client.chat.completions.create(
model=self.model_name,
messages=dialogue,
stream=True
logger.bind(tag=TAG).debug(
f"Sending request to Xinference with model: {self.model_name}, dialogue length: {len(dialogue)}"
)
is_active=True
responses = self.client.chat.completions.create(
model=self.model_name, messages=dialogue, stream=True
)
is_active = True
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 ''
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:
if "<think>" in content:
is_active = False
content = content.split('<think>')[0]
if '</think>' in content:
content = content.split("<think>")[0]
if "</think>" in content:
is_active = True
content = content.split('</think>')[-1]
content = content.split("</think>")[-1]
if is_active:
yield content
except Exception as e:
@@ -59,10 +65,14 @@ class LLMProvider(LLMProviderBase):
def response_with_functions(self, session_id, dialogue, functions=None):
try:
logger.bind(tag=TAG).debug(f"Sending function call request to Xinference with model: {self.model_name}, dialogue length: {len(dialogue)}")
logger.bind(tag=TAG).debug(
f"Sending function call request to Xinference with model: {self.model_name}, dialogue length: {len(dialogue)}"
)
if functions:
logger.bind(tag=TAG).debug(f"Function calls enabled with: {[f.get('function', {}).get('name') for f in functions]}")
logger.bind(tag=TAG).debug(
f"Function calls enabled with: {[f.get('function', {}).get('name') for f in functions]}"
)
stream = self.client.chat.completions.create(
model=self.model_name,
messages=dialogue,
@@ -74,7 +84,7 @@ class LLMProvider(LLMProviderBase):
delta = chunk.choices[0].delta
content = delta.content
tool_calls = delta.tool_calls
if content:
yield content, tool_calls
elif tool_calls:
@@ -82,4 +92,7 @@ class LLMProvider(LLMProviderBase):
except Exception as e:
logger.bind(tag=TAG).error(f"Error in Xinference function call: {e}")
yield {"type": "content", "content": f"【Xinference服务响应异常: {str(e)}】"}
yield {
"type": "content",
"content": f"【Xinference服务响应异常: {str(e)}】",
}
main/xiaozhi-server/core/providers/memory/mem_local_short/mem_local_short.py
+10 -2
View File
@@ -167,7 +167,12 @@ class MemoryProvider(MemoryProviderBase):
msgStr += f"当前时间：{time_str}"
if self.save_to_file:
result = self.llm.response_no_stream(short_term_memory_prompt, msgStr)
result = self.llm.response_no_stream(
short_term_memory_prompt,
msgStr,
max_tokens=2000,
temperature=0.2,
)
json_str = extract_json_data(result)
try:
json.loads(json_str) # 检查json格式是否正确
@@ -177,7 +182,10 @@ class MemoryProvider(MemoryProviderBase):
print("Error:", e)
else:
result = self.llm.response_no_stream(
short_term_memory_prompt_only_content, msgStr
short_term_memory_prompt_only_content,
msgStr,
max_tokens=2000,
temperature=0.2,
)
save_mem_local_short(self.role_id, result)
logger.bind(tag=TAG).info(f"Save memory successful - Role: {self.role_id}")