Merge branch 'main' into tts-response

main/xiaozhi-server/core/providers/llm/base.py

@@ -10,7 +10,7 @@ class LLMProviderBase(ABC):
         """LLM response generator"""
         pass
 
-    def response_no_stream(self, system_prompt, user_prompt):
+    def response_no_stream(self, system_prompt, user_prompt, **kwargs):
         try:
             # 构造对话格式
             dialogue = [
@@ -18,7 +18,7 @@ class LLMProviderBase(ABC):
                 {"role": "user", "content": user_prompt}
             ]
             result = ""
-            for part in self.response("", dialogue):
+            for part in self.response("", dialogue, **kwargs):
                 result += part
             return result
 
@@ -30,7 +30,7 @@ class LLMProviderBase(ABC):
         """
         Default implementation for function calling (streaming)
         This should be overridden by providers that support function calls
-        
+
         Returns: generator that yields either text tokens or a special function call token
         """
         # For providers that don't support functions, just return regular response

main/xiaozhi-server/core/providers/llm/coze/coze.py

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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

@@ -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/openai/openai.py

@@ -16,26 +16,37 @@ class LLMProvider(LLMProviderBase):
             self.base_url = config.get("base_url")
         else:
             self.base_url = config.get("url")
-        max_tokens = config.get("max_tokens")
-        if max_tokens is None or max_tokens == "":
-            max_tokens = 500
 
-        try:
-            max_tokens = int(max_tokens)
-        except (ValueError, TypeError):
-            max_tokens = 500
-        self.max_tokens = max_tokens
+        param_defaults = {
+            "max_tokens": (500, int),
+            "temperature": (0.7, lambda x: round(float(x), 1)),
+            "top_p": (1.0, lambda x: round(float(x), 1)),
+            "frequency_penalty": (0, lambda x: round(float(x), 1))
+        }
+
+        for param, (default, converter) in param_defaults.items():
+            value = config.get(param)
+            try:
+                setattr(self, param, converter(value) if value not in (None, "") else default)
+            except (ValueError, TypeError):
+                setattr(self, param, default)
+
+        logger.debug(
+            f"意图识别参数初始化: {self.temperature}, {self.max_tokens}, {self.top_p}, {self.frequency_penalty}")
 
         check_model_key("LLM", self.api_key)
         self.client = openai.OpenAI(api_key=self.api_key, base_url=self.base_url)
 
-    def response(self, session_id, dialogue):
+    def response(self, session_id, dialogue, **kwargs):
         try:
             responses = self.client.chat.completions.create(
                 model=self.model_name,
                 messages=dialogue,
                 stream=True,
-                max_tokens=self.max_tokens,
+                max_tokens=kwargs.get("max_tokens", self.max_tokens),
+                temperature=kwargs.get("temperature", self.temperature),
+                top_p=kwargs.get("top_p", self.top_p),
+                frequency_penalty=kwargs.get("frequency_penalty", self.frequency_penalty),
             )
 
             is_active = True

main/xiaozhi-server/core/providers/llm/xinference/xinference.py

@@ -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/base.py

@@ -9,7 +9,13 @@ class MemoryProviderBase(ABC):
     def __init__(self, config):
         self.config = config
         self.role_id = None
-        self.llm = None
+
+    def set_llm(self, llm):
+        self.llm = llm
+        # 获取模型名称和类型信息
+        model_name = getattr(llm, "model_name", str(llm.__class__.__name__))
+        # 记录更详细的日志
+        logger.bind(tag=TAG).info(f"记忆总结设置LLM: {model_name}")
 
     @abstractmethod
     async def save_memory(self, msgs):

main/xiaozhi-server/core/providers/memory/mem_local_short/mem_local_short.py

@@ -107,7 +107,7 @@ TAG = __name__
 class MemoryProvider(MemoryProviderBase):
     def __init__(self, config, summary_memory):
         super().__init__(config)
-        self.short_momery = ""
+        self.short_memory = ""
         self.save_to_file = True
         self.memory_path = get_project_dir() + "data/.memory.yaml"
         self.load_memory(summary_memory)
@@ -122,7 +122,7 @@ class MemoryProvider(MemoryProviderBase):
     def load_memory(self, summary_memory):
         # api获取到总结记忆后直接返回
         if summary_memory or not self.save_to_file:
-            self.short_momery = summary_memory
+            self.short_memory = summary_memory
             return
 
         all_memory = {}
@@ -130,18 +130,21 @@ class MemoryProvider(MemoryProviderBase):
             with open(self.memory_path, "r", encoding="utf-8") as f:
                 all_memory = yaml.safe_load(f) or {}
         if self.role_id in all_memory:
-            self.short_momery = all_memory[self.role_id]
+            self.short_memory = all_memory[self.role_id]
 
     def save_memory_to_file(self):
         all_memory = {}
         if os.path.exists(self.memory_path):
             with open(self.memory_path, "r", encoding="utf-8") as f:
                 all_memory = yaml.safe_load(f) or {}
-        all_memory[self.role_id] = self.short_momery
+        all_memory[self.role_id] = self.short_memory
         with open(self.memory_path, "w", encoding="utf-8") as f:
             yaml.dump(all_memory, f, allow_unicode=True)
 
     async def save_memory(self, msgs):
+        # 打印使用的模型信息
+        model_info = getattr(self.llm, "model_name", str(self.llm.__class__.__name__))
+        logger.bind(tag=TAG).debug(f"使用记忆保存模型: {model_info}")
         if self.llm is None:
             logger.bind(tag=TAG).error("LLM is not set for memory provider")
             return None
@@ -155,31 +158,39 @@ class MemoryProvider(MemoryProviderBase):
                 msgStr += f"User: {msg.content}\n"
             elif msg.role == "assistant":
                 msgStr += f"Assistant: {msg.content}\n"
-        if self.short_momery and len(self.short_momery) > 0:
+        if self.short_memory and len(self.short_memory) > 0:
             msgStr += "历史记忆：\n"
-            msgStr += self.short_momery
+            msgStr += self.short_memory
 
         # 当前时间
         time_str = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
         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
-                self.short_momery = json_str
+                self.short_memory = json_str
                 self.save_memory_to_file()
             except Exception as e:
                 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}")
 
-        return self.short_momery
+        return self.short_memory
 
     async def query_memory(self, query: str) -> str:
-        return self.short_momery
+        return self.short_memory

main/xiaozhi-server/core/providers/tts/aliyun.py

@@ -91,11 +91,24 @@ class TTSProvider(TTSProviderBase):
 
         self.appkey = config.get("appkey")
         self.format = config.get("format", "wav")
-        self.sample_rate = config.get("sample_rate", 16000)
-        self.voice = config.get("voice", "xiaoyun")
-        self.volume = config.get("volume", 50)
-        self.speech_rate = config.get("speech_rate", 0)
-        self.pitch_rate = config.get("pitch_rate", 0)
+
+        sample_rate = config.get("sample_rate", "16000")
+        self.sample_rate = int(sample_rate) if sample_rate else 16000
+
+        if config.get("private_voice"):
+            self.voice = config.get("private_voice")
+        else:
+            self.voice = config.get("voice", "xiaoyun")
+
+        volume = config.get("volume", "50")
+        self.volume = int(volume) if volume else 50
+
+        speech_rate = config.get("speech_rate", "0")
+        self.speech_rate = int(speech_rate) if speech_rate else 0
+
+        pitch_rate = config.get("pitch_rate", "0")
+        self.pitch_rate = int(pitch_rate) if pitch_rate else 0
+
         self.host = config.get("host", "nls-gateway-cn-shanghai.aliyuncs.com")
         self.api_url = f"https://{self.host}/stream/v1/tts"
         self.header = {"Content-Type": "application/json"}