Merge pull request #3270 from xinnan-tech/fix-voiceprint-dialogue-bug

fix： 修改提示词与说话人注入逻辑，减少模型回答带着说话人信息频率

main/xiaozhi-server/agent-base-prompt.txt

@@ -65,7 +65,7 @@ Calling unnecessarily is also wrong: User asks ”明天天气” → you call g
 
 <speaker_recognition>
 For the input format `{"speaker":"...", "content":"..."}` (speaker = speaker name, content = text):
-1. [Identity known] When `speaker` is a specific name, identity has been recognized. On the first turn you must address them naturally and adjust your response style based on their history.
+1. [Identity known] When `speaker` is a specific name, identity has been recognized — remember who they are. Do not start every reply by addressing them by name; get straight to the content (a natural greeting by name once, at the very start of a conversation, is fine). If they ask about their identity (e.g. "你知道我是谁吗", "还记得我吗"), answer truthfully using their recognized name.
 2. [Identity unknown] When `speaker` is `未知说话人`, the system failed to identify the voice. You must NEVER mention the data inside the `speakers_info` tag to the user. Judge from context whether the speaker is the owner or the owner's friend, and keep the conversation natural.
 </speaker_recognition>
 

main/xiaozhi-server/core/connection.py

@@ -159,6 +159,8 @@ class ConnectionHandler:
         self.asr_audio = []  # 存储PCM帧列表，供VAD和ASR共享
         self.asr_audio_queue = queue.Queue()
         self.current_speaker = None  # 存储当前说话人
+        self.introduced_speakers = set()  # 已"首次引入"的说话人，控制只在首轮带名字
+        self.system_introduced_speakers = set()  # 已在 system 注入过身份的说话人，控制 system 身份只首轮出现
 
         # llm相关变量
         self.dialogue = Dialogue()
@@ -1120,12 +1122,20 @@ class ConnectionHandler:
                 )
                 memory_str = future.result()
 
+            # 仅在该说话人首次出现时把身份注入 system，之后靠对话历史首轮保留，
+            # 避免每轮在 system 重复出现名字诱导模型反复称呼
+            speaker_for_system = None
+            cs = (self.current_speaker or "").strip()
+            if cs and cs != "未知说话人" and cs not in self.system_introduced_speakers:
+                self.system_introduced_speakers.add(cs)
+                speaker_for_system = cs
+
             if self.intent_type == "function_call" and functions is not None:
                 # 使用支持functions的streaming接口
                 llm_responses = self.llm.response_with_functions(
                     self.session_id,
                     self.dialogue.get_llm_dialogue_with_memory(
-                        memory_str, self.config.get("voiceprint", {})
+                        memory_str, self.config.get("voiceprint", {}), speaker_for_system
                     ),
                     functions=functions,
                 )
@@ -1133,7 +1143,7 @@ class ConnectionHandler:
                 llm_responses = self.llm.response(
                     self.session_id,
                     self.dialogue.get_llm_dialogue_with_memory(
-                        memory_str, self.config.get("voiceprint", {})
+                        memory_str, self.config.get("voiceprint", {}), speaker_for_system
                     ),
                 )
         except Exception as e:

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

@@ -43,7 +43,6 @@ async def resume_vad_detection(conn: "ConnectionHandler"):
 async def startToChat(conn: "ConnectionHandler", text):
     # 检查输入是否是JSON格式（包含说话人信息）
     speaker_name = None
-    language_tag = None
     actual_text = text
 
     try:
@@ -52,12 +51,16 @@ async def startToChat(conn: "ConnectionHandler", text):
             data = json.loads(text)
             if "speaker" in data and "content" in data:
                 speaker_name = data["speaker"]
-                language_tag = data["language"]
-                actual_text = data["content"]
+                actual_content = data["content"]
                 conn.logger.bind(tag=TAG).info(f"解析到说话人信息: {speaker_name}")
 
-                # 直接使用JSON格式的文本，不解析
-                actual_text = text
+                # 仅在该说话人首次出现时保留 {"speaker":...} JSON，让模型自然称呼一次；
+                # 后续轮降为纯文本，避免每轮重复出现名字诱导模型反复称呼
+                if speaker_name not in conn.introduced_speakers:
+                    conn.introduced_speakers.add(speaker_name)
+                    actual_text = text
+                else:
+                    actual_text = actual_content
     except (json.JSONDecodeError, KeyError):
         # 如果解析失败，继续使用原始文本
         pass

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

@@ -43,11 +43,8 @@ class MemoryProvider(MemoryProviderBase):
         self.memory_client = None
         self.enable_user_profile = False
         self.last_profile_content = ""  # Cache for user profile from UserMemory
-
         try:
-            # Check if user profile mode is enabled
-            self.enable_user_profile = config.get("enable_user_profile", False)
-            
+            self.enable_user_profile = str(config.get("enable_user_profile", False)).lower() == 'true'
             # Get configuration parameters
             database_provider = config.get("database_provider", "sqlite")
             llm_provider = config.get("llm_provider", "qwen")

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

@@ -92,7 +92,8 @@ class Dialogue:
         return result
 
     def get_llm_dialogue_with_memory(
-            self, memory_str: str = None, voiceprint_config: dict = None
+            self, memory_str: str = None, voiceprint_config: dict = None,
+            current_speaker: str = None,
     ) -> List[Dict[str, str]]:
         # 构建对话
         dialogue = []
@@ -103,50 +104,31 @@ class Dialogue:
         )
 
         if system_message:
-            # 以 <context> 为分界点，拆分静态 system prompt 和动态上下文
-            # 静态部分（规则、身份等）保持不变，可命中前缀缓存
-            # 动态部分（时间、天气、记忆等）作为第二条 system 消息，保持 system 权威性
             full_prompt = system_message.content
-            context_match = re.search(r"<context>", full_prompt)
-            if context_match:
-                static_part = full_prompt[:context_match.start()]
-                dynamic_part = full_prompt[context_match.start():]
-            else:
-                static_part = full_prompt
-                dynamic_part = ""
 
-            # 第一段：静态 system prompt（前缀缓存可命中）
-            dialogue.append({"role": "system", "content": static_part})
-
-        # 第二段：few-shot 示例（会话内不变，也是缓存前缀的一部分）
-        non_system_messages = [m for m in self.dialogue if m.role != "system"]
-        fewshot_messages = [m for m in non_system_messages if m.is_temporary]
-        complete_fewshot = self._ensure_tool_calls_complete(fewshot_messages)
-        for m in complete_fewshot:
-            self.getMessages(m, dialogue)
-
-        # 第三段：动态上下文 system prompt（时间、记忆、说话人等）
-        # 保持 system 角色以确保模型权威性，不降级为 user
-        if system_message and dynamic_part:
             # 替换时间占位符
-            dynamic_part = dynamic_part.replace(
+            full_prompt = full_prompt.replace(
                 "{{current_time}}", datetime.now().strftime("%H:%M")
             )
 
             # 填充记忆
             if memory_str is not None:
-                dynamic_part = re.sub(
+                full_prompt = re.sub(
                     r"<memory>.*?</memory>",
                     f"<memory>\n{memory_str}\n</memory>",
-                    dynamic_part,
+                    full_prompt,
                     flags=re.DOTALL,
                 )
 
             # 追加说话人信息
             try:
-                speakers = voiceprint_config.get("speakers", [])
-                if speakers:
-                    dynamic_part += "\n<speakers_info>"
+                current_speaker_name = (current_speaker or "").strip()
+                # 仅在本轮注入了有效身份时才输出 speakers_info，避免列表里的名字每轮
+                # 重复出现诱导模型反复称呼；后续轮不再注入身份，靠对话历史首轮保留
+                if current_speaker_name and current_speaker_name != "未知说话人":
+                    speakers = voiceprint_config.get("speakers", [])
+                    speakers_info = "\n<speakers_info>"
+                    speakers_info += f"\n当前说话人：{current_speaker_name}"
                     for speaker_str in speakers:
                         try:
                             parts = speaker_str.split(",", 2)
@@ -155,16 +137,24 @@ class Dialogue:
                                 description = (
                                     parts[2].strip() if len(parts) >= 3 else ""
                                 )
-                                dynamic_part += f"\n- {name}：{description}"
+                                speakers_info += f"\n- {name}：{description}"
                         except:
                             pass
-                    dynamic_part += "\n</speakers_info>"
+                    speakers_info += "\n</speakers_info>"
+                    full_prompt += speakers_info
             except:
                 pass
 
-            dialogue.append({"role": "system", "content": dynamic_part})
+            dialogue.append({"role": "system", "content": full_prompt})
 
-        # 第四段：实际对话历史（不含 few-shot）
+        # 第二段：few-shot 示例（会话内不变）
+        non_system_messages = [m for m in self.dialogue if m.role != "system"]
+        fewshot_messages = [m for m in non_system_messages if m.is_temporary]
+        complete_fewshot = self._ensure_tool_calls_complete(fewshot_messages)
+        for m in complete_fewshot:
+            self.getMessages(m, dialogue)
+
+        # 第三段：实际对话历史（不含 few-shot）
         actual_messages = [m for m in non_system_messages if not m.is_temporary]
         complete_actual = self._ensure_tool_calls_complete(actual_messages)
         for m in complete_actual: