feat(memory): unify alias extraction into metadata pipeline and deduplicate user entity nodes

- Merge alias add/remove into MetadataExtractionResponse and Celery metadata task,
  removing the separate sync step from extraction_orchestrator
- Replace first-person pronouns ("我") with "用户" in statement extraction to
  preserve identity semantics for downstream metadata/alias extraction
- Update extract_statement.jinja2 prompt to enforce "用户" as subject for user
  statements instead of resolving to real names
- Add alias change instructions (aliases_to_add/aliases_to_remove) to
  extract_user_metadata.jinja2 with incremental merge logic
- Deduplicate special entities ("用户", "AI助手") in graph_saver by reusing
  existing Neo4j node IDs per end_user_id
- Sync final aliases from PgSQL to Neo4j user entity nodes after metadata write

api/app/core/memory/models/metadata_models.py

@@ -38,3 +38,11 @@ class MetadataExtractionResponse(BaseModel):
     """元数据提取 LLM 响应结构"""
     model_config = ConfigDict(extra='ignore')
     user_metadata: UserMetadata = Field(default_factory=UserMetadata)
+    aliases_to_add: List[str] = Field(
+        default_factory=list,
+        description="本次新发现的用户别名（用户自我介绍或他人对用户的称呼）"
+    )
+    aliases_to_remove: List[str] = Field(
+        default_factory=list,
+        description="用户明确否认的别名（如'我不叫XX了'）"
+    )

api/app/core/memory/storage_services/extraction_engine/extraction_orchestrator.py

@@ -311,9 +311,8 @@ class ExtractionOrchestrator:
                 dialog_data_list,
             )
 
-            # 步骤 7: 同步用户别名到数据库表 + 触发异步元数据提取（仅正式模式）
+            # 步骤 7: 触发异步元数据和别名提取（仅正式模式）
             if not is_pilot_run:
-                # 收集用户相关 statement 并触发异步元数据提取
                 try:
                     from app.core.memory.storage_services.extraction_engine.knowledge_extraction.metadata_extractor import MetadataExtractor
                     metadata_extractor = MetadataExtractor(llm_client=self.llm_client, language=self.language)
@@ -322,7 +321,6 @@ class ExtractionOrchestrator:
                         statement_entity_edges
                     )
                     if user_statements:
-                        # 获取 end_user_id 和 config_id
                         end_user_id = dialog_data_list[0].end_user_id if dialog_data_list else None
                         config_id = dialog_data_list[0].config_id if dialog_data_list and hasattr(dialog_data_list[0], 'config_id') else None
                         if end_user_id:
@@ -339,9 +337,7 @@ class ExtractionOrchestrator:
                 except Exception as e:
                     logger.error(f"触发元数据提取任务失败（不影响主流程）: {e}", exc_info=True)
 
-                # 同步用户别名到数据库表
-                logger.info("步骤 7: 同步用户别名到 end_user 和 end_user_info 表")
-                await self._update_end_user_other_name(entity_nodes, dialog_data_list)
+                # 别名同步已迁移到 Celery 元数据提取任务中，不再在此处执行
 
             logger.info(f"知识提取流水线运行完成（{mode_str}）")
             return (

api/app/core/memory/storage_services/extraction_engine/knowledge_extraction/metadata_extractor.py

@@ -14,9 +14,9 @@ from app.core.memory.models.graph_models import (
     StatementNode,
 )
 from app.core.memory.models.metadata_models import (
-    MetadataExtractionResponse,
     UserMetadata,
 )
+from app.core.memory.models.message_models import DialogData
 
 logger = logging.getLogger(__name__)
 
@@ -50,6 +50,35 @@ class MetadataExtractor:
             return "zh"
         return "en"
 
+    # def collect_user_raw_messages(
+    #     self,
+    #     dialog_data_list: List[DialogData],
+    # ) -> List[str]:
+    #     """
+    #     从原始对话数据中提取 speaker="user" 的消息原文。
+
+    #     直接使用用户的原始输入，不经过陈述句提取阶段的 LLM 改写，
+    #     避免第一人称被替换为第三人称导致元数据/别名提取错误。
+
+    #     Returns:
+    #         用户原始消息文本列表
+    #     """
+    #     result = []
+    #     for dialog in dialog_data_list:
+    #         if not dialog.context or not dialog.context.msgs:
+    #             continue
+    #         for msg in dialog.context.msgs:
+    #             if getattr(msg, 'role', '') == 'user':
+    #                 text = (getattr(msg, 'msg', '') or '').strip()
+    #                 if text:
+    #                     result.append(text)
+
+    #     logger.info(f"收集到 {len(result)} 条用户原始消息")
+    #     if result:
+    #         for i, text in enumerate(result):
+    #             logger.info(f"  [user message {i+1}] {text[:200]}")
+    #     return result
+
     def collect_user_related_statements(
         self,
         entity_nodes: List[ExtractedEntityNode],
@@ -104,6 +133,9 @@ class MetadataExtractor:
             f"(直接关联: {total_associated}, speaker=user: {len(result)}, "
             f"跳过非user: {skipped_non_user})"
         )
+        if result:
+            for i, text in enumerate(result):
+                logger.info(f"  [user statement {i+1}] {text}")
         if total_associated > 0 and len(result) == 0:
             logger.warning(
                 f"有 {total_associated} 条直接关联 statement 但全部被 speaker 过滤，"
@@ -111,18 +143,22 @@ class MetadataExtractor:
             )
         return result
 
-    async def extract_metadata(self, statements: List[str], existing_metadata: Optional[dict] = None) -> Optional[UserMetadata]:
+    async def extract_metadata(
+        self,
+        statements: List[str],
+        existing_metadata: Optional[dict] = None,
+        existing_aliases: Optional[List[str]] = None,
+    ) -> Optional[tuple]:
         """
-        对筛选后的 statement 列表调用 LLM 提取元数据。
-        语言根据 statement 内容自动检测，不依赖系统界面语言。
-        传入已有元数据作为上下文，让 LLM 能判断 replace/remove 操作。
+        对筛选后的 statement 列表调用 LLM 提取元数据和用户别名。
 
         Args:
             statements: 用户发言的 statement 文本列表
-            existing_metadata: 数据库已有的元数据（可选），用于 LLM 对比判断变更
+            existing_metadata: 数据库已有的元数据（可选）
+            existing_aliases: 数据库已有的用户别名列表（可选）
 
         Returns:
-            UserMetadata on success, None on failure
+            (UserMetadata, List[str], List[str]) tuple: (metadata, aliases_to_add, aliases_to_remove) on success, None on failure
         """
         if not statements:
             return None
@@ -130,7 +166,6 @@ class MetadataExtractor:
         try:
             from app.core.memory.utils.prompt.prompt_utils import prompt_env
 
-            # 根据写入内容的语言自动检测，而非使用系统界面语言
             detected_language = self.detect_language(statements)
             logger.info(f"元数据提取语言检测结果: {detected_language}")
 
@@ -139,18 +174,23 @@ class MetadataExtractor:
                 statements=statements,
                 language=detected_language,
                 existing_metadata=existing_metadata,
+                existing_aliases=existing_aliases,
                 json_schema="",
             )
 
+            from app.core.memory.models.metadata_models import MetadataExtractionResponse
             response = await self.llm_client.response_structured(
                 messages=[{"role": "user", "content": prompt}],
                 response_model=MetadataExtractionResponse,
             )
 
-            if response and response.user_metadata:
-                return response.user_metadata
+            if response:
+                metadata = response.user_metadata if response.user_metadata else None
+                to_add = response.aliases_to_add if response.aliases_to_add else []
+                to_remove = response.aliases_to_remove if response.aliases_to_remove else []
+                return metadata, to_add, to_remove
 
-            logger.warning("LLM 返回的元数据为空")
+            logger.warning("LLM 返回的响应为空")
             return None
 
         except Exception as e:

api/app/core/memory/storage_services/extraction_engine/knowledge_extraction/statement_extraction.py

@@ -82,6 +82,18 @@ class StatementExtractor:
         logger.warning(f"Chunk {getattr(chunk, 'id', 'unknown')} has no speaker field or is empty")
         return None
 
+    @staticmethod
+    def _replace_first_person_with_user(text: str) -> str:
+        """将用户消息中的第一人称代词"我"替换为"用户"。
+        
+        替换规则：
+        - 所有独立的"我"都替换为"用户"（包括"我的"→"用户的"、"叫我"→"叫用户"）
+        - 不替换"我们"中的"我"（"我们"是复数，不指代用户个人）
+        """
+        import re
+        result = re.sub(r'我(?!们)', '用户', text)
+        return result
+
     async def _extract_statements(self, chunk, end_user_id: Optional[str] = None, dialogue_content: str = None) -> List[Statement]:
         """Process a single chunk and return extracted statements
 
@@ -99,6 +111,13 @@ class StatementExtractor:
             logger.warning(f"Chunk {chunk.id} content too short or empty, skipping")
             return []
 
+        # 对 speaker="user" 的 chunk，将第一人称"我"替换为"用户"，
+        # 避免 LLM 在陈述句提取时将"我"替换为具体名字（如"齐齐"），
+        # 导致下游元数据/别名提取无法识别第一人称语义。
+        chunk_speaker = self._get_speaker_from_chunk(chunk)
+        if chunk_speaker == "user":
+            chunk_content = self._replace_first_person_with_user(chunk_content)
+
         prompt_content = await render_statement_extraction_prompt(
             chunk_content=chunk_content,
             definitions=LABEL_DEFINITIONS,
@@ -149,8 +168,6 @@ class StatementExtractor:
                     relevence_info = RelevenceInfo[relevence_str] if relevence_str in RelevenceInfo.__members__ else RelevenceInfo.RELEVANT
                 except (KeyError, ValueError):
                     relevence_info = RelevenceInfo.RELEVANT
-               
-                chunk_speaker = self._get_speaker_from_chunk(chunk)
             
                 chunk_statement = Statement(
                     statement=extracted_stmt.statement,

api/app/core/memory/utils/prompt/prompts/extract_statement.jinja2

@@ -43,8 +43,9 @@ Each statement must be labeled as per the criteria mentioned below.
 
 对话上下文和共指消解：
 - 将每个陈述句归属于说出它的参与者。
-- 如果参与者列表为说话者提供了名称（例如，"李雪（用户）"），请在提取的陈述句中使用具体名称（"李雪"），而不是通用角色（"用户"）。
-- 将所有代词解析为对话上下文中的具体人物或实体。
+- **对于用户的发言：必须使用"用户"作为主语**，禁止将"用户"或"我"替换为用户的真实姓名或别名。例如，用户说"我叫张三"应提取为"用户叫张三"，而不是"张三叫张三"。
+- 对于 AI 助手的发言：使用"助手"或"AI助手"作为主语。
+- 将所有代词解析为对话上下文中的具体人物或实体，但"我"必须解析为"用户"。
 - 识别并将抽象引用解析为其具体名称（如果提到）。
 - 将缩写和首字母缩略词扩展为其完整形式。
 {% else %}
@@ -68,8 +69,9 @@ Context Resolution Requirements:
 
 Conversational Context & Co-reference Resolution:
 - Attribute every statement to the participant who uttered it.
-- If the participant list provides a name for a speaker (e.g., "李雪 (用户)"), use the specific name ("李雪") in the extracted statement, not the generic role ("用户").
-- Resolve all pronouns to the specific person or entity from the conversation's context.
+- **For user's statements: always use "用户" (User) as the subject**. Do NOT replace "用户" or "I" with the user's real name or alias. For example, if the user says "I'm John", extract as "用户 is John", not "John is John".
+- For AI assistant's statements: use "助手" or "AI助手" as the subject.
+- Resolve all pronouns to the specific person or entity from the conversation's context, but "I"/"我" must always resolve to "用户".
 - Identify and resolve abstract references to their specific names if mentioned.
 - Expand abbreviations and acronyms to their full form.
 {% endif %}
@@ -139,13 +141,13 @@ AI: "水彩画很有趣！水彩颜料通常由颜料与阿拉伯树胶等粘合
 示例输出: {
   "statements": [
     {
-      "statement": "Sarah Chen 最近一直在尝试水彩画。",
+      "statement": "用户最近一直在尝试水彩画。",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "Sarah Chen 画了一些花朵。",
+      "statement": "用户画了一些花朵。",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
@@ -157,13 +159,13 @@ AI: "水彩画很有趣！水彩颜料通常由颜料与阿拉伯树胶等粘合
       "relevance": "IRRELEVANT"
     },
     {
-      "statement": "Sarah Chen 认为她的水彩画中的色彩组合可以改进。",
+      "statement": "用户认为她的水彩画中的色彩组合可以改进。",
       "statement_type": "OPINION",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "Sarah Chen 真的很喜欢玫瑰和百合。",
+      "statement": "用户真的很喜欢玫瑰和百合。",
       "statement_type": "FACT",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
@@ -186,13 +188,13 @@ AI: "水彩画很有趣！水彩颜料通常由颜料和阿拉伯树胶等粘合
 示例输出: {
   "statements": [
     {
-      "statement": "张曼婷最近在尝试水彩画。",
+      "statement": "用户最近在尝试水彩画。",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "张曼婷画了一些花朵。",
+      "statement": "用户画了一些花朵。",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
@@ -204,13 +206,13 @@ AI: "水彩画很有趣！水彩颜料通常由颜料和阿拉伯树胶等粘合
       "relevance": "IRRELEVANT"
     },
     {
-      "statement": "张曼婷觉得水彩画的色彩搭配还有提升的空间。",
+      "statement": "用户觉得水彩画的色彩搭配还有提升的空间。",
       "statement_type": "OPINION",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "张曼婷很喜欢玫瑰和百合。",
+      "statement": "用户很喜欢玫瑰和百合。",
       "statement_type": "FACT",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
@@ -233,13 +235,13 @@ User: "I think the color combinations could use some improvement, but I really l
 Example Output: {
   "statements": [
     {
-      "statement": "Sarah Chen has been trying watercolor painting recently.",
+      "statement": "用户 has been trying watercolor painting recently.",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "Sarah Chen painted some flowers.",
+      "statement": "用户 painted some flowers.",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
@@ -251,13 +253,13 @@ Example Output: {
       "relevance": "IRRELEVANT"
     },
     {
-      "statement": "Sarah Chen thinks the color combinations in her watercolor paintings could use some improvement.",
+      "statement": "用户 thinks the color combinations in her watercolor paintings could use some improvement.",
       "statement_type": "OPINION",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "Sarah Chen really likes roses and lilies.",
+      "statement": "用户 really likes roses and lilies.",
       "statement_type": "FACT",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
@@ -280,13 +282,13 @@ AI: "水彩画很有趣！水彩颜料通常由颜料和阿拉伯树胶等粘合
 Example Output: {
   "statements": [
     {
-      "statement": "张曼婷最近在尝试水彩画。",
+      "statement": "用户最近在尝试水彩画。",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "张曼婷画了一些花朵。",
+      "statement": "用户画了一些花朵。",
       "statement_type": "FACT",
       "temporal_type": "DYNAMIC",
       "relevance": "RELEVANT"
@@ -298,13 +300,13 @@ Example Output: {
       "relevance": "IRRELEVANT"
     },
     {
-      "statement": "张曼婷觉得水彩画的色彩搭配还有提升的空间。",
+      "statement": "用户觉得水彩画的色彩搭配还有提升的空间。",
       "statement_type": "OPINION",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"
     },
     {
-      "statement": "张曼婷很喜欢玫瑰和百合。",
+      "statement": "用户很喜欢玫瑰和百合。",
       "statement_type": "FACT",
       "temporal_type": "STATIC",
       "relevance": "RELEVANT"

api/app/core/memory/utils/prompt/prompts/extract_user_metadata.jinja2

@@ -32,6 +32,22 @@ Extract user metadata from the following conversation statements spoken by the u
 - behavioral_hints.preferred_depth：偏好深度（概览/技术细节/深入探讨）
 - behavioral_hints.tone_preference：语气偏好（轻松随意/专业简洁/学术严谨）
 - knowledge_tags：用户涉及的知识领域标签
+
+**用户别名变更（增量模式）：**
+- **aliases_to_add**：本次新发现的用户别名，包括：
+  * 用户主动自我介绍：如"我叫张三"、"我的名字是XX"、"我的网名是XX"
+  * 他人对用户的称呼：如"同事叫我陈哥"、"大家叫我小张"、"领导叫我老陈"
+  * 只提取原文中逐字出现的名字，严禁推测或创造
+  * 禁止提取：用户给 AI 取的名字、第三方人物自身的名字、"用户"/"我" 等占位词
+  * 如果没有新别名，返回空数组 `[]`
+- **aliases_to_remove**：用户明确否认的别名，包括：
+  * 用户说"我不叫XX了"、"别叫我XX"、"我改名了，不叫XX" → 将 XX 放入此数组
+  * **严格限制**：只将用户原文中**逐字提到**的被否认名字放入，不要推断关联的其他别名
+  * 例如：用户说"我不叫陈小刀了" → 只移除"陈小刀"，不要移除"陈哥"、"老陈"等未被提及的别名
+  * 如果没有要移除的别名，返回空数组 `[]`
+{% if existing_aliases %}
+- 已有别名：{{ existing_aliases | tojson }}（仅供参考，不需要在输出中重复）
+{% endif %}
 {% else %}
 **"Three-Degree Principle" criteria:**
 - Reusability: Will this information be used by multiple functional modules?
@@ -63,6 +79,22 @@ Existing user metadata from the database is provided below. Combine with the use
 - behavioral_hints.preferred_depth: Preferred depth (overview/detailed/deep dive)
 - behavioral_hints.tone_preference: Tone preference (casual/professional/academic)
 - knowledge_tags: Knowledge domain tags related to the user
+
+**User alias changes (incremental mode):**
+- **aliases_to_add**: Newly discovered user aliases from this conversation, including:
+  * User self-introductions: e.g. "I'm John", "My name is XX", "My username is XX"
+  * How others address the user: e.g. "My colleagues call me Johnny", "People call me Mike"
+  * Only extract names that appear VERBATIM in the text — never infer or fabricate
+  * Do NOT extract: names the user gives to the AI, third-party people's own names, placeholder words like "User"/"I"
+  * If no new aliases, return empty array `[]`
+- **aliases_to_remove**: Aliases the user explicitly denies, including:
+  * User says "Don't call me XX anymore", "I'm not called XX", "I changed my name from XX" → put XX in this array
+  * **Strict rule**: Only include the exact name the user **verbatim mentions** as denied. Do NOT infer or remove related aliases
+  * Example: User says "I'm not called John anymore" → only remove "John", do NOT remove "Johnny", "J" or other related aliases not mentioned
+  * If no aliases to remove, return empty array `[]`
+{% if existing_aliases %}
+- Existing aliases: {{ existing_aliases | tojson }} (for reference only, do not repeat in output)
+{% endif %}
 {% endif %}
 
 ===User Statements===
@@ -94,7 +126,9 @@ Return a JSON object with the following structure:
       "tone_preference": ""
     },
     "knowledge_tags": []
-  }
+  },
+  "aliases_to_add": [],
+  "aliases_to_remove": []
 }
 ```