update：使用intent_llm中的LLM进行工具内容回复

main/xiaozhi-server/config.yaml

@@ -157,7 +157,7 @@ selected_module:
   Memory: nomem
   # 意图识别模块开启后，可以播放音乐、控制音量、识别退出指令。
   # 不想开通意图识别，就设置成：nointent
-  # 意图识别可使用intent_llm。优点：通用性强，缺点：增加串行前置意图识别模块，会增加处理时间，这个意图识别暂时不支持控制音量大小等iot操作
+  # 意图识别可使用intent_llm。优点：通用性强，缺点：增加串行前置意图识别模块，会增加处理时间，支持控制音量大小等iot操作
   # 意图识别可使用function_call，缺点：需要所选择的LLM支持function_call，优点：按需调用工具、速度快，理论上能全部操作所有iot指令
   # 默认免费的ChatGLMLLM就已经支持function_call，但是如果像追求稳定建议把LLM设置成：DoubaoLLM，使用的具体model_name是：doubao-1-5-pro-32k-250115
   Intent: function_call

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

@@ -108,43 +108,21 @@ async def process_intent_result(conn, intent_result, original_text):
                     if result.action == Action.RESPONSE:  # 直接回复前端
                         text = result.response
                         if text is not None:
-                            text_index = (
-                                conn.tts_last_text_index + 1
-                                if hasattr(conn, "tts_last_text_index")
-                                else 0
-                            )
-                            conn.recode_first_last_text(text, text_index)
-                            future = conn.executor.submit(
-                                conn.speak_and_play, text, text_index
-                            )
-                            conn.llm_finish_task = True
-                            conn.tts_queue.put((future, text_index))
-                            conn.dialogue.put(Message(role="assistant", content=text))
+                            speak_and_play(conn, 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.dialogue.put(Message(role="tool", content=text))
+                        llm_result = conn.intent.replyResult(text)
+                        if llm_result is None:
+                            llm_result = text
+                        speak_and_play(conn, llm_result)
                     elif (
                         result.action == Action.NOTFOUND
                         or result.action == Action.ERROR
                     ):
                         text = result.result
                         if text is not None:
-                            text_index = (
-                                conn.tts_last_text_index + 1
-                                if hasattr(conn, "tts_last_text_index")
-                                else 0
-                            )
-                            conn.recode_first_last_text(text, text_index)
-                            future = conn.executor.submit(
-                                conn.speak_and_play, text, text_index
-                            )
-                            conn.llm_finish_task = True
-                            conn.tts_queue.put((future, text_index))
-                            conn.dialogue.put(Message(role="assistant", content=text))
+                            speak_and_play(conn, text)
                     elif function_name != "play_music":
                         # For backward compatibility with original code
                         # 获取当前最新的文本索引
@@ -152,18 +130,7 @@ async def process_intent_result(conn, intent_result, original_text):
                         if text is None:
                             text = result.result
                         if text is not None:
-                            text_index = (
-                                conn.tts_last_text_index + 1
-                                if hasattr(conn, "tts_last_text_index")
-                                else 0
-                            )
-                            conn.recode_first_last_text(text, text_index)
-                            future = conn.executor.submit(
-                                conn.speak_and_play, text, text_index
-                            )
-                            conn.llm_finish_task = True
-                            conn.tts_queue.put((future, text_index))
-                            conn.dialogue.put(Message(role="assistant", content=text))
+                            speak_and_play(conn, text)
 
             # 将函数执行放在线程池中
             conn.executor.submit(process_function_call)
@@ -174,21 +141,12 @@ async def process_intent_result(conn, intent_result, original_text):
         return False
 
 
-def extract_text_in_brackets(s):
-    """
-    从字符串中提取中括号内的文字
-
-    :param s: 输入字符串
-    :return: 中括号内的文字，如果不存在则返回空字符串
-    """
-    left_bracket_index = s.find("[")
-    right_bracket_index = s.find("]")
-
-    if (
-        left_bracket_index != -1
-        and right_bracket_index != -1
-        and left_bracket_index < right_bracket_index
-    ):
-        return s[left_bracket_index + 1 : right_bracket_index]
-    else:
-        return ""
+def speak_and_play(conn, text):
+    text_index = (
+        conn.tts_last_text_index + 1 if hasattr(conn, "tts_last_text_index") else 0
+    )
+    conn.recode_first_last_text(text, text_index)
+    future = conn.executor.submit(conn.speak_and_play, text, text_index)
+    conn.llm_finish_task = True
+    conn.tts_queue.put((future, text_index))
+    conn.dialogue.put(Message(role="assistant", content=text))

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

@@ -20,6 +20,7 @@ class IntentProvider(IntentProviderBase):
         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
+        self.history_count = 4  # 默认使用最近4条对话记录
 
     def get_intent_system_prompt(self, functions_list: str) -> str:
         """
@@ -104,6 +105,13 @@ class IntentProvider(IntentProviderBase):
             for key, _ in sorted_items[: len(sorted_items) - self.cache_max_size]:
                 del self.intent_cache[key]
 
+    def replyResult(self, text: str):
+        llm_result = self.llm.response_no_stream(
+            system_prompt=text,
+            user_prompt="请总结以上内容，像人类一样说话的口吻回复用户，要求简洁，请直接返回结果。",
+        )
+        return llm_result
+
     async def detect_intent(self, conn, dialogue_history: List[Dict], text: str) -> str:
         if not self.llm:
             raise ValueError("LLM provider not set")
@@ -150,14 +158,13 @@ class IntentProvider(IntentProviderBase):
 
         logger.bind(tag=TAG).debug(f"User prompt: {prompt_music}")
 
-        # 构建用户最后一句话的提示
+        # 构建用户对话历史的提示
         msgStr = ""
 
-        # 只使用最后两句即可
-        if len(dialogue_history) >= 2:
-            # 保证最少有两句话的时候处理
-            msgStr += f"{dialogue_history[-2].role}: {dialogue_history[-2].content}\n"
-        msgStr += f"{dialogue_history[-1].role}: {dialogue_history[-1].content}\n"
+        # 获取最近的对话历史
+        start_idx = max(0, len(dialogue_history) - self.history_count)
+        for i in range(start_idx, len(dialogue_history)):
+            msgStr += f"{dialogue_history[i].role}: {dialogue_history[i].content}\n"
 
         msgStr += f"User: {text}\n"
         user_prompt = f"current dialogue:\n{msgStr}"