update:intent_llm加载动态插件

main/xiaozhi-server/config.yaml

@@ -175,6 +175,13 @@ Intent:
     # 如果这里不填，则会默认使用selected_module.LLM的模型作为意图识别的思考模型
     # 如果你的不想使用selected_module.LLM意图识别，这里最好使用独立的LLM作为意图识别，例如使用免费的ChatGLMLLM
     llm: ChatGLMLLM
+    # plugins_func/functions下的模块，可以通过配置，选择加载哪个模块，加载后对话支持相应的function调用
+    # 系统默认已经记载“handle_exit_intent(退出识别)”、“play_music(音乐播放)”插件，请勿重复加载
+    # 下面是加载查天气、角色切换、加载查新闻的插件示例
+    functions:
+      - get_weather
+      - get_news_from_newsnow
+      - play_music
   function_call:
     # 不需要动type
     type: function_call

main/xiaozhi-server/core/connection.py

@@ -58,9 +58,9 @@ class ConnectionHandler:
         self.config = copy.deepcopy(config)
         self.session_id = str(uuid.uuid4())
         self.logger = setup_logging()
-        self.auth = AuthMiddleware(config)
         self.server = server  # 保存server实例的引用
 
+        self.auth = AuthMiddleware(config)
         self.need_bind = False
         self.bind_code = None
         self.read_config_from_api = self.config.get("read_config_from_api", False)
@@ -128,8 +128,10 @@ class ConnectionHandler:
             if len(cmd) > self.max_cmd_length:
                 self.max_cmd_length = len(cmd)
 
-        self.close_after_chat = False  # 是否在聊天结束后关闭连接
+        # 是否在聊天结束后关闭连接
-        self.use_function_call_mode = False
+        self.close_after_chat = False
+        self.load_function_plugin = False
+        self.intent_type = "nointent"
 
         self.timeout_task = None
         self.timeout_seconds = (
@@ -365,11 +367,11 @@ class ConnectionHandler:
         self.memory.init_memory(self.device_id, self.llm)
 
     def _initialize_intent(self):
-        if (
+        self.intent_type = self.config["Intent"][
-            self.config["Intent"][self.config["selected_module"]["Intent"]]["type"]
+            self.config["selected_module"]["Intent"]
-            == "function_call"
+        ]["type"]
-        ):
+        if self.intent_type == "function_call" or self.intent_type == "intent_llm":
-            self.use_function_call_mode = True
+            self.load_function_plugin = True
         """初始化意图识别模块"""
         # 获取意图识别配置
         intent_config = self.config["Intent"]
@@ -757,7 +759,13 @@ class ConnectionHandler:
                 )
 
                 self.dialogue.put(
-                    Message(role="tool", tool_call_id=function_id, content=text)
+                    Message(
+                        role="tool",
+                        tool_call_id=(
+                            str(uuid.uuid4()) if function_id is None else function_id
+                        ),
+                        content=text,
+                    )
                 )
                 self.chat_with_function_calling(text, tool_call=True)
         elif result.action == Action.NOTFOUND or result.action == Action.ERROR:

main/xiaozhi-server/core/handle/intentHandler.py

@@ -19,7 +19,7 @@ async def handle_user_intent(conn, text):
     if await checkWakeupWords(conn, text):
         return True
 
-    if conn.use_function_call_mode:
+    if conn.intent_type == "function_call":
         # 使用支持function calling的聊天方法,不再进行意图分析
         return False
     # 使用LLM进行意图分析
@@ -103,7 +103,7 @@ async def process_intent_result(conn, intent_result, original_text):
                     conn, function_call_data
                 )
                 logger.bind(tag=TAG).debug(f"检测到Action : {result.action}")
-                
+
                 if result:
                     if result.action == Action.RESPONSE:  # 直接回复前端
                         text = result.response
@@ -118,14 +118,19 @@ async def process_intent_result(conn, intent_result, original_text):
                                 conn.speak_and_play, text, text_index
                             )
                             conn.llm_finish_task = True
-                            conn.tts_queue.put(future)
+                            conn.tts_queue.put((future, text_index))
                             conn.dialogue.put(Message(role="assistant", content=text))
                     elif result.action == Action.REQLLM:  # 调用函数后再请求llm生成回复
                         text = result.result
                         if text is not None and len(text) > 0:
                             conn.dialogue.put(Message(role="tool", content=text))
-                            conn.executor.submit(conn.chat_with_function_calling, text, True)
+                            conn.executor.submit(
-                    elif result.action == Action.NOTFOUND or result.action == Action.ERROR:
+                                conn.chat_with_function_calling, text, True
+                            )
+                    elif (
+                        result.action == Action.NOTFOUND
+                        or result.action == Action.ERROR
+                    ):
                         text = result.result
                         if text is not None:
                             text_index = (
@@ -157,7 +162,7 @@ async def process_intent_result(conn, intent_result, original_text):
                                 conn.speak_and_play, text, text_index
                             )
                             conn.llm_finish_task = True
-                            conn.tts_queue.put(future)
+                            conn.tts_queue.put((future, text_index))
                             conn.dialogue.put(Message(role="assistant", content=text))
 
             # 将函数执行放在线程池中

main/xiaozhi-server/core/handle/iotHandle.py

@@ -317,7 +317,7 @@ async def handleIotDescriptors(conn, descriptors):
         )
         conn.iot_descriptors[descriptor["name"]] = iot_descriptor
 
-        if conn.use_function_call_mode:
+        if conn.load_function_plugin:
             # 注册或获取设备类型
             type_id = register_device_type(descriptor)
             device_functions = device_type_registry.get_device_functions(type_id)

main/xiaozhi-server/core/handle/receiveAudioHandle.py

@@ -76,7 +76,7 @@ async def startToChat(conn, text):
 
     # 意图未被处理，继续常规聊天流程
     await send_stt_message(conn, text)
-    if conn.use_function_call_mode:
+    if conn.intent_type == "function_call":
         # 使用支持function calling的聊天方法
         conn.executor.submit(conn.chat_with_function_calling, text)
     else:

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

@@ -9,18 +9,6 @@ logger = setup_logging()
 class IntentProviderBase(ABC):
     def __init__(self, config):
         self.config = config
-        self.intent_options = [
-            {
-                "name": "handle_exit_intent",
-                "desc": "结束聊天, 用户发来如再见之类的表示结束的话, 不想再进行对话的时候",
-            },
-            {
-                "name": "play_music",
-                "desc": "播放音乐, 用户希望你可以播放音乐, 只用于播放音乐的意图",
-            },
-            {"name": "get_time", "desc": "获取今天日期或者当前时间信息"},
-            {"name": "continue_chat", "desc": "继续聊天"},
-        ]
 
     def set_llm(self, llm):
         self.llm = llm

main/xiaozhi-server/core/providers/intent/intent_llm/intent_llm.py

@@ -15,57 +15,71 @@ class IntentProvider(IntentProviderBase):
     def __init__(self, config):
         super().__init__(config)
         self.llm = None
-        self.promot = self.get_intent_system_prompt()
+        self.promot = ""
         # 添加缓存管理
         self.intent_cache = {}  # 缓存意图识别结果
         self.cache_expiry = 600  # 缓存有效期10分钟
         self.cache_max_size = 100  # 最多缓存100个意图
 
-    def get_intent_system_prompt(self) -> str:
+    def get_intent_system_prompt(self, functions_list: str) -> str:
         """
-        根据配置的意图选项动态生成系统提示词
+        根据配置的意图选项和可用函数动态生成系统提示词
+        Args:
+            functions: 可用的函数列表，JSON格式字符串
         Returns:
             格式化后的系统提示词
         """
 
+        # 构建函数说明部分
+        functions_desc = "可用的函数列表：\n"
+        for func in functions_list:
+            func_info = func.get("function", {})
+            name = func_info.get("name", "")
+            desc = func_info.get("description", "")
+            params = func_info.get("parameters", {})
+
+            functions_desc += f"\n函数名: {name}\n"
+            functions_desc += f"描述: {desc}\n"
+
+            if params:
+                functions_desc += "参数:\n"
+                for param_name, param_info in params.get("properties", {}).items():
+                    param_desc = param_info.get("description", "")
+                    param_type = param_info.get("type", "")
+                    functions_desc += f"- {param_name} ({param_type}): {param_desc}\n"
+
+            functions_desc += "---\n"
+
         prompt = (
-            "你是一个意图识别助手。请分析用户的最后一句话，判断用户意图属于以下哪一类：\n"
+            "你是一个意图识别助手。请分析用户的最后一句话，判断用户意图并调用相应的函数。\n\n"
-            "<start>"
+            f"{functions_desc}\n"
-            f"{str(self.intent_options)}"
+            "处理步骤:\n"
-            "<end>\n"
+            "1. 分析用户输入，确定用户意图\n"
-            "处理步骤:"
+            "2. 从可用函数列表中选择最匹配的函数\n"
-            "1. 思考意图类型，生成function_call格式"
+            "3. 如果找到匹配的函数，生成对应的function_call 格式\n"
-            "\n\n"
+            '4. 如果没有找到匹配的函数，返回{"function_call": {"name": "continue_chat"}}\n\n'
-            "返回格式示例：\n"
+            "返回格式要求：\n"
-            '1. 播放音乐意图: {"function_call": {"name": "play_music", "arguments": {"song_name": "音乐名称"}}}\n'
+            "1. 必须返回纯JSON格式\n"
-            '2. 结束对话意图: {"function_call": {"name": "handle_exit_intent", "arguments": {"say_goodbye": "goodbye"}}}\n'
+            "2. 必须包含function_call字段\n"
-            '3. 获取当天日期时间: {"function_call": {"name": "get_time"}}\n'
+            "3. function_call必须包含name字段\n"
-            '4. 继续聊天意图: {"function_call": {"name": "continue_chat"}}\n'
+            "4. 如果函数需要参数，必须包含arguments字段\n\n"
-            "\n"
+            "示例：\n"
-            "注意:\n"
-            '- 播放音乐：无歌名时，song_name设为"random"\n'
-            "- 如果没有明显的意图，应按照继续聊天意图处理\n"
-            "- 只返回纯JSON，不要任何其他内容\n"
-            "\n"
-            "示例分析:\n"
             "```\n"
-            "用户: 你也太搞笑了\n"
+            "用户: 现在几点了？\n"
-            '返回: {"function_call": {"name": "continue_chat"}}\n'
-            "```\n"
-            "```\n"
-            "用户: 现在是几号了?现在几点了？\n"
             '返回: {"function_call": {"name": "get_time"}}\n'
             "```\n"
             "```\n"
-            "用户: 我们明天再聊吧\n"
+            "用户: 我想结束对话\n"
-            '返回: {"function_call": {"name": "handle_exit_intent"}}\n'
+            '返回: {"function_call": {"name": "handle_exit_intent", "arguments": {"say_goodbye": "goodbye"}}}\n'
             "```\n"
             "```\n"
-            "用户: 播放中秋月\n"
+            "用户: 你好啊\n"
-            '返回: {"function_call": {"name": "play_music", "arguments": {"song_name": "中秋月"}}}\n'
+            '返回: {"function_call": {"name": "continue_chat"}}\n'
-            "```\n"
+            "```\n\n"
-            "```\n"
+            "注意：\n"
-            "可用的音乐名称:\n"
+            "1. 只返回JSON格式，不要包含任何其他文字\n"
+            '2. 如果没有找到匹配的函数，返回{"function_call": {"name": "continue_chat"}}\n'
+            "3. 确保返回的JSON格式正确，包含所有必要的字段\n"
         )
         return prompt
 
@@ -118,6 +132,24 @@ class IntentProvider(IntentProviderBase):
         # 清理缓存
         self.clean_cache()
 
+        if self.promot == "":
+            if hasattr(conn, "func_handler"):
+                functions = conn.func_handler.get_functions()
+                self.promot = self.get_intent_system_prompt(functions)
+
+        music_config = initialize_music_handler(conn)
+        music_file_names = music_config["music_file_names"]
+        prompt_music = f"{self.promot}\n<musicNames>{music_file_names}\n</musicNames>"
+
+        devices = conn.config["plugins"]["home_assistant"].get("devices", [])
+        if len(devices) > 0:
+            hass_prompt = "\n下面是我家智能设备列表（位置，设备名，entity_id），可以通过homeassistant控制\n"
+            for device in devices:
+                hass_prompt += device + "\n"
+            prompt_music += hass_prompt
+
+        logger.bind(tag=TAG).debug(f"User prompt: {prompt_music}")
+
         # 构建用户最后一句话的提示
         msgStr = ""
 
@@ -128,11 +160,7 @@ class IntentProvider(IntentProviderBase):
         msgStr += f"{dialogue_history[-1].role}: {dialogue_history[-1].content}\n"
 
         msgStr += f"User: {text}\n"
-        user_prompt = f"当前的对话如下：\n{msgStr}"
+        user_prompt = f"current dialogue:\n{msgStr}"
-        music_config = initialize_music_handler(conn)
-        music_file_names = music_config["music_file_names"]
-        prompt_music = f"{self.promot}\n<start>{music_file_names}\n<end>"
-        logger.bind(tag=TAG).debug(f"User prompt: {prompt_music}")
 
         # 记录预处理完成时间
         preprocess_time = time.time() - total_start_time

main/xiaozhi-server/core/utils/dialogue.py

@@ -33,7 +33,13 @@ class Dialogue:
             dialogue.append({"role": m.role, "tool_calls": m.tool_calls})
         elif m.role == "tool":
             dialogue.append(
-                {"role": m.role, "tool_call_id": m.tool_call_id, "content": m.content}
+                {
+                    "role": m.role,
+                    "tool_call_id": (
+                        str(uuid.uuid4()) if m.tool_call_id is None else m.tool_call_id
+                    ),
+                    "content": m.content,
+                }
             )
         else:
             dialogue.append({"role": m.role, "content": m.content})

main/xiaozhi-server/plugins_func/functions/hass_init.py

@@ -8,7 +8,7 @@ HASS_CACHE = {}
 
 
 def append_devices_to_prompt(conn):
-    if conn.use_function_call_mode:
+    if conn.intent_type == "function_call":
         funcs = conn.config["Intent"][conn.config["selected_module"]["Intent"]].get(
             "functions", []
         )
@@ -27,7 +27,7 @@ def append_devices_to_prompt(conn):
 def initialize_hass_handler(conn):
     global HASS_CACHE
     if HASS_CACHE == {}:
-        if conn.use_function_call_mode:
+        if conn.load_function_plugin:
             funcs = conn.config["Intent"][conn.config["selected_module"]["Intent"]].get(
                 "functions", []
             )