Phase 4: Agent Assistant + Provenance + Entity Cards + Confidence Indicators

- Add llm_client.py for Kimi API integration with RAG and streaming support
- Add Agent API endpoints: query, command, suggest
- Add Provenance API for relation source tracking
- Add Entity details API with mentions and relations
- Add Entity evolution analysis API
- Update workbench.html with Agent panel, entity cards, provenance modal
- Update app.js with Agent chat, entity hover cards, relation provenance
- Add low-confidence entity highlighting
- Update STATUS.md with Phase 4 progress

STATUS.md

@@ -1,134 +1,100 @@
 # InsightFlow 开发状态
 
-**最后更新**: 2026-02-18
+**最后更新**: 2026-02-19
 
 ## 当前阶段
 
-Phase 3: 记忆与生长 - **已完成 ✅**
+Phase 4: Agent 助手与知识溯源 - **开发中 🚧**
 
 ## 已完成
 
-### Phase 1: 骨架与单体分析 (MVP) ✅
+### Phase 1-3 (已完成 ✅)
+- FastAPI 项目框架搭建
+- SQLite 数据库设计
+- 阿里云听悟 ASR 集成
+- OSS 上传模块
+- 实体提取与对齐逻辑
+- 关系提取
+- 项目 CRUD API
+- 音频上传与分析 API
+- D3.js 知识图谱可视化
+- 实体列表展示
+- 转录文本中实体高亮显示
+- 图谱与文本联动
 
-#### 后端 (backend/)
-- ✅ FastAPI 项目框架搭建
-- ✅ SQLite 数据库设计 (schema.sql)
-- ✅ 数据库管理模块 (db_manager.py)
-- ✅ 阿里云听悟 ASR 集成 (tingwu_client.py)
-- ✅ OSS 上传模块 (oss_uploader.py)
-- ✅ 实体提取与对齐逻辑
-- ✅ 关系提取（LLM 同时提取实体和关系）
-- ✅ 项目 CRUD API
-- ✅ 音频上传与分析 API
-- ✅ 实体列表 API
-- ✅ 关系列表 API
-- ✅ 转录列表 API
-- ✅ 实体提及位置 API
-- ✅ transcripts 表数据写入
-- ✅ entity_mentions 表数据写入
-- ✅ entity_relations 表数据写入
+### Phase 4 - Agent 助手 (已完成 ✅)
+- ✅ 创建 llm_client.py - Kimi API 客户端
+  - 支持流式/非流式聊天
+  - 带置信度的实体提取
+  - RAG 问答功能
+  - Agent 指令解析
+  - 实体演变分析
+- ✅ 更新 db_manager.py - 新增方法
+  - `get_relation_with_details()` - 获取关系详情
+  - `get_entity_with_mentions()` - 获取实体及提及
+  - `search_entities()` - 搜索实体
+  - `update_entity()` - 更新实体
+  - `get_project_summary()` - 项目摘要
+  - `get_transcript_context()` - 转录上下文
+- ✅ 更新 main.py - Agent API 端点
+  - `POST /api/v1/projects/{id}/agent/query` - RAG 问答
+  - `POST /api/v1/projects/{id}/agent/command` - 指令执行
+  - `GET /api/v1/projects/{id}/agent/suggest` - 智能建议
+  - `GET /api/v1/relations/{id}/provenance` - 关系溯源
+  - `GET /api/v1/entities/{id}/details` - 实体详情
+  - `GET /api/v1/entities/{id}/evolution` - 实体演变分析
+  - `GET /api/v1/projects/{id}/entities/search` - 实体搜索
+  - `PATCH /api/v1/entities/{id}` - 更新实体
+- ✅ 更新 workbench.html - Agent 面板 UI
+  - 可折叠的 Agent 助手面板
+  - 聊天界面
+  - 实体悬停卡片
+  - 关系溯源弹窗
+- ✅ 更新 app.js - 前端功能
+  - Agent 聊天功能
+  - 指令执行（合并实体、编辑定义）
+  - RAG 问答
+  - 实体卡片悬停显示
+  - 关系点击溯源
+  - 低置信度实体标黄
 
-#### 前端 (frontend/)
-- ✅ 项目管理页面 (index.html)
-- ✅ 知识工作台页面 (workbench.html)
-- ✅ D3.js 知识图谱可视化
-- ✅ 音频上传 UI
-- ✅ 实体列表展示
-- ✅ 转录文本中实体高亮显示
-- ✅ 图谱与文本联动（点击实体双向高亮）
+### Phase 4 - 知识溯源 (已完成 ✅)
+- ✅ 点击关系连线显示来源文档
+- ✅ 实体详情显示所有提及位置
+- ✅ 证据文本展示
 
-### Phase 2: 交互与纠错工作台 ✅
+### Phase 4 - 术语卡片悬停 (已完成 ✅)
+- ✅ 鼠标悬停实体显示卡片
+- ✅ 卡片包含：名称、定义、提及次数、关系数
 
-#### 后端 API 新增
-- ✅ 实体编辑 API (PUT /api/v1/entities/{id})
-- ✅ 实体删除 API (DELETE /api/v1/entities/{id})
-- ✅ 实体合并 API (POST /api/v1/entities/{id}/merge)
-- ✅ 手动创建实体 API (POST /api/v1/projects/{id}/entities)
-- ✅ 关系创建 API (POST /api/v1/projects/{id}/relations)
-- ✅ 关系删除 API (DELETE /api/v1/relations/{id})
-- ✅ 转录编辑 API (PUT /api/v1/transcripts/{id})
+### Phase 4 - 置信度提示 (已完成 ✅)
+- ✅ LLM 提取返回置信度分数
+- ✅ 低置信度实体在文本中标黄
 
-#### 前端交互功能
-- ✅ 实体编辑器模态框（名称、类型、定义、别名）
-- ✅ 右键菜单（编辑实体、合并实体、标记为实体）
-- ✅ 实体合并功能
-- ✅ 关系管理（添加、删除）
-- ✅ 转录文本编辑模式
-- ✅ 划词创建实体
-- ✅ 文本与图谱双向联动
+## 待完成
 
-#### 数据库更新
-- ✅ update_entity() - 更新实体信息
-- ✅ delete_entity() - 删除实体及关联数据
-- ✅ delete_relation() - 删除关系
-- ✅ update_relation() - 更新关系
-- ✅ update_transcript() - 更新转录文本
-
-### Phase 3: 记忆与生长 ✅
-
-#### 多文件图谱融合
-- ✅ 支持上传多个音频文件到同一项目
-- ✅ 系统自动对齐实体，合并图谱
-- ✅ 实体提及跨文件追踪
-- ✅ 文件选择器切换不同转录内容
-- ✅ 转录列表 API 返回文件类型
-
-#### 实体对齐算法优化
-- ✅ 新增 `entity_aligner.py` 模块
-- ✅ 使用 Kimi API embedding 进行语义相似度匹配
-- ✅ 余弦相似度计算
-- ✅ 自动别名建议
-- ✅ 批量实体对齐 API
-- ✅ 实体对齐回退机制（字符串匹配）
-
-#### PDF/DOCX 文档导入
-- ✅ 新增 `document_processor.py` 模块
-- ✅ 支持 PDF、DOCX、TXT、MD 格式
-- ✅ 文档文本提取并参与实体提取
-- ✅ 文档上传 API (/api/v1/projects/{id}/upload-document)
-- ✅ 文档类型标记（audio/document）
-
-#### 项目知识库面板
-- ✅ 全新的知识库视图
-- ✅ 侧边栏导航切换（工作台/知识库）
-- ✅ 统计面板：实体数、关系数、文件数、术语数
-- ✅ 实体网格展示（带提及统计）
-- ✅ 关系列表展示
-- ✅ 术语表管理（添加/删除）
-- ✅ 文件列表展示（区分音频/文档）
-
-#### 术语表功能
-- ✅ 术语表数据库表 (glossary)
-- ✅ 添加术语 API
-- ✅ 获取术语列表 API
-- ✅ 删除术语 API
-- ✅ 前端术语表管理界面
-
-#### 数据库更新
-- ✅ transcripts 表新增 `type` 字段
-- ✅ entities 表新增 `embedding` 字段
-- ✅ 新增 glossary 表
-- ✅ 新增索引优化查询性能
+### Phase 4 - Neo4j 集成 (可选)
+- [ ] 将图谱数据同步到 Neo4j
+- [ ] 支持复杂图查询
 
 ## 技术债务
 
 - 听悟 SDK fallback 到 mock 需要更好的错误处理
+- 实体相似度匹配目前只是简单字符串包含，需要 embedding 方案
 - 前端需要状态管理（目前使用全局变量）
 - 需要添加 API 文档 (OpenAPI/Swagger)
-- Embedding 缓存需要持久化
-- 实体对齐算法需要更多测试
 
 ## 部署信息
 
 - 服务器: 122.51.127.111
 - 项目路径: /opt/projects/insightflow
 - 端口: 18000
-- Docker 镜像: insightflow:phase3
-- 最后部署: 2026-02-19 06:05 AM
 
-## 下一步 (Phase 4)
+## 最近更新
 
-- 知识推理与问答
-- 实体属性扩展
-- 时间线视图
-- 导出功能（PDF/图片）
+### 2026-02-19
+- 完成 Phase 4 Agent 助手功能
+- 实现知识溯源功能
+- 添加术语卡片悬停
+- 实现置信度提示
+- 更新前端 UI 和交互

backend/db_manager.py

@@ -454,6 +454,168 @@ class DatabaseManager:
         """获取所有实体用于 embedding 计算"""
         return self.list_project_entities(project_id)
     
+    # Phase 4: Agent & Provenance methods
+    def get_relation_with_details(self, relation_id: str) -> Optional[dict]:
+        """获取关系详情，包含源文档信息"""
+        conn = self.get_conn()
+        row = conn.execute(
+            """SELECT r.*, 
+                s.name as source_name, t.name as target_name,
+                tr.filename as transcript_filename, tr.full_text as transcript_text
+               FROM entity_relations r
+               JOIN entities s ON r.source_entity_id = s.id
+               JOIN entities t ON r.target_entity_id = t.id
+               LEFT JOIN transcripts tr ON r.transcript_id = tr.id
+               WHERE r.id = ?""",
+            (relation_id,)
+        ).fetchone()
+        conn.close()
+        if row:
+            return dict(row)
+        return None
+    
+    def get_entity_with_mentions(self, entity_id: str) -> Optional[dict]:
+        """获取实体详情及所有提及位置"""
+        conn = self.get_conn()
+        
+        # 获取实体信息
+        entity_row = conn.execute(
+            "SELECT * FROM entities WHERE id = ?", (entity_id,)
+        ).fetchone()
+        
+        if not entity_row:
+            conn.close()
+            return None
+        
+        entity = dict(entity_row)
+        entity['aliases'] = json.loads(entity['aliases']) if entity['aliases'] else []
+        
+        # 获取提及位置
+        mentions = conn.execute(
+            """SELECT m.*, t.filename, t.created_at as transcript_date
+               FROM entity_mentions m
+               JOIN transcripts t ON m.transcript_id = t.id
+               WHERE m.entity_id = ?
+               ORDER BY t.created_at, m.start_pos""",
+            (entity_id,)
+        ).fetchall()
+        
+        entity['mentions'] = [dict(m) for m in mentions]
+        entity['mention_count'] = len(mentions)
+        
+        # 获取相关关系
+        relations = conn.execute(
+            """SELECT r.*, 
+                s.name as source_name, t.name as target_name
+               FROM entity_relations r
+               JOIN entities s ON r.source_entity_id = s.id
+               JOIN entities t ON r.target_entity_id = t.id
+               WHERE r.source_entity_id = ? OR r.target_entity_id = ?
+               ORDER BY r.created_at DESC""",
+            (entity_id, entity_id)
+        ).fetchall()
+        
+        entity['relations'] = [dict(r) for r in relations]
+        
+        conn.close()
+        return entity
+    
+    def search_entities(self, project_id: str, query: str) -> List[Entity]:
+        """搜索实体"""
+        conn = self.get_conn()
+        rows = conn.execute(
+            """SELECT * FROM entities 
+               WHERE project_id = ? AND 
+               (name LIKE ? OR definition LIKE ? OR aliases LIKE ?)
+               ORDER BY name""",
+            (project_id, f'%{query}%', f'%{query}%', f'%{query}%')
+        ).fetchall()
+        conn.close()
+        
+        entities = []
+        for row in rows:
+            data = dict(row)
+            data['aliases'] = json.loads(data['aliases']) if data['aliases'] else []
+            entities.append(Entity(**data))
+        return entities
+    
+    def get_project_summary(self, project_id: str) -> dict:
+        """获取项目摘要信息，用于 RAG 上下文"""
+        conn = self.get_conn()
+        
+        # 项目基本信息
+        project = conn.execute(
+            "SELECT * FROM projects WHERE id = ?", (project_id,)
+        ).fetchone()
+        
+        # 统计信息
+        entity_count = conn.execute(
+            "SELECT COUNT(*) as count FROM entities WHERE project_id = ?",
+            (project_id,)
+        ).fetchone()['count']
+        
+        transcript_count = conn.execute(
+            "SELECT COUNT(*) as count FROM transcripts WHERE project_id = ?",
+            (project_id,)
+        ).fetchone()['count']
+        
+        relation_count = conn.execute(
+            "SELECT COUNT(*) as count FROM entity_relations WHERE project_id = ?",
+            (project_id,)
+        ).fetchone()['count']
+        
+        # 获取最近的转录文本片段
+        recent_transcripts = conn.execute(
+            """SELECT filename, full_text, created_at 
+               FROM transcripts 
+               WHERE project_id = ? 
+               ORDER BY created_at DESC 
+               LIMIT 5""",
+            (project_id,)
+        ).fetchall()
+        
+        # 获取高频实体
+        top_entities = conn.execute(
+            """SELECT e.name, e.type, e.definition, COUNT(m.id) as mention_count
+               FROM entities e
+               LEFT JOIN entity_mentions m ON e.id = m.entity_id
+               WHERE e.project_id = ?
+               GROUP BY e.id
+               ORDER BY mention_count DESC
+               LIMIT 10""",
+            (project_id,)
+        ).fetchall()
+        
+        conn.close()
+        
+        return {
+            'project': dict(project) if project else {},
+            'statistics': {
+                'entity_count': entity_count,
+                'transcript_count': transcript_count,
+                'relation_count': relation_count
+            },
+            'recent_transcripts': [dict(t) for t in recent_transcripts],
+            'top_entities': [dict(e) for e in top_entities]
+        }
+    
+    def get_transcript_context(self, transcript_id: str, position: int, context_chars: int = 200) -> str:
+        """获取转录文本的上下文"""
+        conn = self.get_conn()
+        row = conn.execute(
+            "SELECT full_text FROM transcripts WHERE id = ?",
+            (transcript_id,)
+        ).fetchone()
+        conn.close()
+        
+        if not row:
+            return ""
+        
+        text = row['full_text']
+        start = max(0, position - context_chars)
+        end = min(len(text), position + context_chars)
+        return text[start:end]
+
 
 # Singleton instance
 _db_manager = None

backend/llm_client.py

@@ -0,0 +1,255 @@
+#!/usr/bin/env python3
+"""
+InsightFlow LLM Client - Phase 4
+用于与 Kimi API 交互，支持 RAG 问答和 Agent 功能
+"""
+
+import os
+import json
+import httpx
+from typing import List, Dict, Optional, AsyncGenerator
+from dataclasses import dataclass
+
+KIMI_API_KEY = os.getenv("KIMI_API_KEY", "")
+KIMI_BASE_URL = os.getenv("KIMI_BASE_URL", "https://api.kimi.com/coding")
+
+
+@dataclass
+class ChatMessage:
+    role: str
+    content: str
+
+
+@dataclass
+class EntityExtractionResult:
+    name: str
+    type: str
+    definition: str
+    confidence: float
+
+
+@dataclass
+class RelationExtractionResult:
+    source: str
+    target: str
+    type: str
+    confidence: float
+
+
+class LLMClient:
+    """Kimi API 客户端"""
+    
+    def __init__(self, api_key: str = None, base_url: str = None):
+        self.api_key = api_key or KIMI_API_KEY
+        self.base_url = base_url or KIMI_BASE_URL
+        self.headers = {
+            "Authorization": f"Bearer {self.api_key}",
+            "Content-Type": "application/json"
+        }
+    
+    async def chat(self, messages: List[ChatMessage], temperature: float = 0.3, stream: bool = False) -> str:
+        """发送聊天请求"""
+        if not self.api_key:
+            raise ValueError("KIMI_API_KEY not set")
+        
+        payload = {
+            "model": "k2p5",
+            "messages": [{"role": m.role, "content": m.content} for m in messages],
+            "temperature": temperature,
+            "stream": stream
+        }
+        
+        async with httpx.AsyncClient() as client:
+            response = await client.post(
+                f"{self.base_url}/v1/chat/completions",
+                headers=self.headers,
+                json=payload,
+                timeout=120.0
+            )
+            response.raise_for_status()
+            result = response.json()
+            return result["choices"][0]["message"]["content"]
+    
+    async def chat_stream(self, messages: List[ChatMessage], temperature: float = 0.3) -> AsyncGenerator[str, None]:
+        """流式聊天请求"""
+        if not self.api_key:
+            raise ValueError("KIMI_API_KEY not set")
+        
+        payload = {
+            "model": "k2p5",
+            "messages": [{"role": m.role, "content": m.content} for m in messages],
+            "temperature": temperature,
+            "stream": True
+        }
+        
+        async with httpx.AsyncClient() as client:
+            async with client.stream(
+                "POST",
+                f"{self.base_url}/v1/chat/completions",
+                headers=self.headers,
+                json=payload,
+                timeout=120.0
+            ) as response:
+                response.raise_for_status()
+                async for line in response.aiter_lines():
+                    if line.startswith("data: "):
+                        data = line[6:]
+                        if data == "[DONE]":
+                            break
+                        try:
+                            chunk = json.loads(data)
+                            delta = chunk["choices"][0]["delta"]
+                            if "content" in delta:
+                                yield delta["content"]
+                        except:
+                            pass
+    
+    async def extract_entities_with_confidence(self, text: str) -> tuple[List[EntityExtractionResult], List[RelationExtractionResult]]:
+        """提取实体和关系，带置信度分数"""
+        prompt = f"""从以下会议文本中提取关键实体和它们之间的关系，以 JSON 格式返回：
+
+文本：{text[:3000]}
+
+要求：
+1. entities: 每个实体包含 name(名称), type(类型: PROJECT/TECH/PERSON/ORG/OTHER), definition(一句话定义), confidence(置信度0-1)
+2. relations: 每个关系包含 source(源实体名), target(目标实体名), type(关系类型: belongs_to/works_with/depends_on/mentions/related), confidence(置信度0-1)
+3. 只返回 JSON 对象，格式: {{"entities": [...], "relations": [...]}}
+
+示例：
+{{
+  "entities": [
+    {{"name": "Project Alpha", "type": "PROJECT", "definition": "核心项目", "confidence": 0.95}},
+    {{"name": "K8s", "type": "TECH", "definition": "Kubernetes容器编排平台", "confidence": 0.88}}
+  ],
+  "relations": [
+    {{"source": "Project Alpha", "target": "K8s", "type": "depends_on", "confidence": 0.82}}
+  ]
+}}"""
+        
+        messages = [ChatMessage(role="user", content=prompt)]
+        content = await self.chat(messages, temperature=0.1)
+        
+        import re
+        json_match = re.search(r'\{{.*?\}}', content, re.DOTALL)
+        if not json_match:
+            return [], []
+        
+        try:
+            data = json.loads(json_match.group())
+            entities = [
+                EntityExtractionResult(
+                    name=e["name"],
+                    type=e.get("type", "OTHER"),
+                    definition=e.get("definition", ""),
+                    confidence=e.get("confidence", 0.8)
+                )
+                for e in data.get("entities", [])
+            ]
+            relations = [
+                RelationExtractionResult(
+                    source=r["source"],
+                    target=r["target"],
+                    type=r.get("type", "related"),
+                    confidence=r.get("confidence", 0.8)
+                )
+                for r in data.get("relations", [])
+            ]
+            return entities, relations
+        except Exception as e:
+            print(f"Parse extraction result failed: {e}")
+            return [], []
+    
+    async def rag_query(self, query: str, context: str, project_context: Dict) -> str:
+        """RAG 问答 - 基于项目上下文回答问题"""
+        prompt = f"""你是一个专业的项目分析助手。基于以下项目信息回答问题：
+
+## 项目信息
+{json.dumps(project_context, ensure_ascii=False, indent=2)}
+
+## 相关上下文
+{context[:4000]}
+
+## 用户问题
+{query}
+
+请用中文回答，保持简洁专业。如果信息不足，请明确说明。"""
+        
+        messages = [
+            ChatMessage(role="system", content="你是一个专业的项目分析助手，擅长从会议记录中提取洞察。"),
+            ChatMessage(role="user", content=prompt)
+        ]
+        
+        return await self.chat(messages, temperature=0.3)
+    
+    async def agent_command(self, command: str, project_context: Dict) -> Dict:
+        """Agent 指令解析 - 将自然语言指令转换为结构化操作"""
+        prompt = f"""解析以下用户指令，转换为结构化操作：
+
+## 项目信息
+{json.dumps(project_context, ensure_ascii=False, indent=2)}
+
+## 用户指令
+{command}
+
+请分析指令意图，返回 JSON 格式：
+{{
+  "intent": "merge_entities|answer_question|edit_entity|create_relation|unknown",
+  "params": {{
+    // 根据 intent 不同，参数不同
+  }},
+  "explanation": "对用户指令的解释"
+}}
+
+意图说明：
+- merge_entities: 合并实体，params 包含 source_names(源实体名列表), target_name(目标实体名)
+- answer_question: 回答问题，params 包含 question(问题内容)
+- edit_entity: 编辑实体，params 包含 entity_name(实体名), field(字段), value(新值)
+- create_relation: 创建关系，params 包含 source(源实体), target(目标实体), relation_type(关系类型)
+"""
+        
+        messages = [ChatMessage(role="user", content=prompt)]
+        content = await self.chat(messages, temperature=0.1)
+        
+        import re
+        json_match = re.search(r'\{{.*?\}}', content, re.DOTALL)
+        if not json_match:
+            return {"intent": "unknown", "explanation": "无法解析指令"}
+        
+        try:
+            return json.loads(json_match.group())
+        except:
+            return {"intent": "unknown", "explanation": "解析失败"}
+    
+    async def analyze_entity_evolution(self, entity_name: str, mentions: List[Dict]) -> str:
+        """分析实体在项目中的演变/态度变化"""
+        mentions_text = "\n".join([
+            f"[{m.get('created_at', '未知时间')}] {m.get('text_snippet', '')}"
+            for m in mentions[:20]  # 限制数量
+        ])
+        
+        prompt = f"""分析实体 "{entity_name}" 在项目中的演变和态度变化：
+
+## 提及记录
+{mentions_text}
+
+请分析：
+1. 该实体的角色/重要性变化
+2. 相关方对它的态度变化
+3. 关键时间节点
+4. 总结性洞察
+
+用中文回答，结构清晰。"""
+        
+        messages = [ChatMessage(role="user", content=prompt)]
+        return await self.chat(messages, temperature=0.3)
+
+
+# Singleton instance
+_llm_client = None
+
+
+def get_llm_client() -> LLMClient:
+    global _llm_client
+    if _llm_client is None:
+        _llm_client = LLMClient()
+    return _llm_client

backend/main.py

@@ -48,6 +48,12 @@ try:
 except ImportError:
     ALIGNER_AVAILABLE = False
 
+try:
+    from llm_client import get_llm_client, ChatMessage
+    LLM_CLIENT_AVAILABLE = True
+except ImportError:
+    LLM_CLIENT_AVAILABLE = False
+
 app = FastAPI(title="InsightFlow", version="0.3.0")
 
 app.add_middleware(
@@ -99,6 +105,13 @@ class RelationCreate(BaseModel):
 class TranscriptUpdate(BaseModel):
     full_text: str
 
+class AgentQuery(BaseModel):
+    query: str
+    stream: bool = False
+
+class AgentCommand(BaseModel):
+    command: str
+
 class EntityMergeRequest(BaseModel):
     source_entity_id: str
     target_entity_id: str
@@ -963,13 +976,14 @@ async def get_entity_mentions(entity_id: str):
 async def health_check():
     return {
         "status": "ok",
-        "version": "0.3.0",
-        "phase": "Phase 3 - Memory & Growth",
+        "version": "0.4.0",
+        "phase": "Phase 4 - Agent Assistant",
         "oss_available": OSS_AVAILABLE,
         "tingwu_available": TINGWU_AVAILABLE,
         "db_available": DB_AVAILABLE,
         "doc_processor_available": DOC_PROCESSOR_AVAILABLE,
-        "aligner_available": ALIGNER_AVAILABLE
+        "aligner_available": ALIGNER_AVAILABLE,
+        "llm_client_available": LLM_CLIENT_AVAILABLE
     }
 
 # Serve frontend
@@ -978,3 +992,276 @@ app.mount("/", StaticFiles(directory="frontend", html=True), name="frontend")
 if __name__ == "__main__":
     import uvicorn
     uvicorn.run(app, host="0.0.0.0", port=8000)
+
+
+# ==================== Phase 4: Agent 助手 API ====================
+
+@app.post("/api/v1/projects/{project_id}/agent/query")
+async def agent_query(project_id: str, query: AgentQuery):
+    """Agent RAG 问答"""
+    if not DB_AVAILABLE or not LLM_CLIENT_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Service not available")
+    
+    db = get_db_manager()
+    llm = get_llm_client()
+    
+    project = db.get_project(project_id)
+    if not project:
+        raise HTTPException(status_code=404, detail="Project not found")
+    
+    # 获取项目上下文
+    project_context = db.get_project_summary(project_id)
+    
+    # 构建上下文
+    context_parts = []
+    for t in project_context.get('recent_transcripts', []):
+        context_parts.append(f"【{t['filename']}】\n{t['full_text'][:1000]}")
+    
+    context = "\n\n".join(context_parts)
+    
+    if query.stream:
+        from fastapi.responses import StreamingResponse
+        import json
+        
+        async def stream_response():
+            messages = [
+                ChatMessage(role="system", content="你是一个专业的项目分析助手，擅长从会议记录中提取洞察。"),
+                ChatMessage(role="user", content=f"""基于以下项目信息回答问题：
+
+## 项目信息
+{json.dumps(project_context, ensure_ascii=False, indent=2)}
+
+## 相关上下文
+{context[:4000]}
+
+## 用户问题
+{query.query}
+
+请用中文回答，保持简洁专业。如果信息不足，请明确说明。""")
+            ]
+            
+            async for chunk in llm.chat_stream(messages):
+                yield f"data: {json.dumps({'content': chunk})}\n\n"
+            yield "data: [DONE]\n\n"
+        
+        return StreamingResponse(stream_response(), media_type="text/event-stream")
+    else:
+        answer = await llm.rag_query(query.query, context, project_context)
+        return {"answer": answer, "project_id": project_id}
+
+
+@app.post("/api/v1/projects/{project_id}/agent/command")
+async def agent_command(project_id: str, command: AgentCommand):
+    """Agent 指令执行 - 解析并执行自然语言指令"""
+    if not DB_AVAILABLE or not LLM_CLIENT_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Service not available")
+    
+    db = get_db_manager()
+    llm = get_llm_client()
+    
+    project = db.get_project(project_id)
+    if not project:
+        raise HTTPException(status_code=404, detail="Project not found")
+    
+    # 获取项目上下文
+    project_context = db.get_project_summary(project_id)
+    
+    # 解析指令
+    parsed = await llm.agent_command(command.command, project_context)
+    
+    intent = parsed.get("intent", "unknown")
+    params = parsed.get("params", {})
+    
+    result = {"intent": intent, "explanation": parsed.get("explanation", "")}
+    
+    # 执行指令
+    if intent == "merge_entities":
+        # 合并实体
+        source_names = params.get("source_names", [])
+        target_name = params.get("target_name", "")
+        
+        target_entity = None
+        source_entities = []
+        
+        # 查找目标实体
+        for e in project_context.get("top_entities", []):
+            if e["name"] == target_name or target_name in e["name"]:
+                target_entity = db.get_entity_by_name(project_id, e["name"])
+                break
+        
+        # 查找源实体
+        for name in source_names:
+            for e in project_context.get("top_entities", []):
+                if e["name"] == name or name in e["name"]:
+                    ent = db.get_entity_by_name(project_id, e["name"])
+                    if ent and (not target_entity or ent.id != target_entity.id):
+                        source_entities.append(ent)
+                    break
+        
+        merged = []
+        if target_entity:
+            for source in source_entities:
+                try:
+                    db.merge_entities(target_entity.id, source.id)
+                    merged.append(source.name)
+                except Exception as e:
+                    print(f"Merge failed: {e}")
+        
+        result["action"] = "merge_entities"
+        result["target"] = target_entity.name if target_entity else None
+        result["merged"] = merged
+        result["success"] = len(merged) > 0
+        
+    elif intent == "answer_question":
+        # 问答 - 调用 RAG
+        answer = await llm.rag_query(params.get("question", command.command), "", project_context)
+        result["action"] = "answer"
+        result["answer"] = answer
+        
+    elif intent == "edit_entity":
+        # 编辑实体
+        entity_name = params.get("entity_name", "")
+        field = params.get("field", "")
+        value = params.get("value", "")
+        
+        entity = db.get_entity_by_name(project_id, entity_name)
+        if entity:
+            updated = db.update_entity(entity.id, **{field: value})
+            result["action"] = "edit_entity"
+            result["entity"] = {"id": updated.id, "name": updated.name} if updated else None
+            result["success"] = updated is not None
+        else:
+            result["success"] = False
+            result["error"] = "Entity not found"
+    
+    else:
+        result["action"] = "none"
+        result["message"] = "无法理解的指令，请尝试：\n- 合并实体：把所有'客户端'合并到'App'\n- 提问：张总对项目的态度如何？\n- 编辑：修改'K8s'的定义为..."
+    
+    return result
+
+
+@app.get("/api/v1/projects/{project_id}/agent/suggest")
+async def agent_suggest(project_id: str):
+    """获取 Agent 建议 - 基于项目数据提供洞察"""
+    if not DB_AVAILABLE or not LLM_CLIENT_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Service not available")
+    
+    db = get_db_manager()
+    llm = get_llm_client()
+    
+    project_context = db.get_project_summary(project_id)
+    
+    # 生成建议
+    prompt = f"""基于以下项目数据，提供3-5条分析建议：
+
+{json.dumps(project_context, ensure_ascii=False, indent=2)}
+
+请提供：
+1. 数据洞察发现
+2. 建议的操作（如合并相似实体、补充定义等）
+3. 值得关注的关键信息
+
+返回 JSON 格式：{{"suggestions": [{{"type": "insight|action", "title": "...", "description": "..."}}]}}"""
+    
+    messages = [ChatMessage(role="user", content=prompt)]
+    content = await llm.chat(messages, temperature=0.3)
+    
+    import re
+    json_match = re.search(r'\{{.*?\}}', content, re.DOTALL)
+    if json_match:
+        try:
+            data = json.loads(json_match.group())
+            return data
+        except:
+            pass
+    
+    return {"suggestions": []}
+
+
+# ==================== Phase 4: 知识溯源 API ====================
+
+@app.get("/api/v1/relations/{relation_id}/provenance")
+async def get_relation_provenance(relation_id: str):
+    """获取关系的知识溯源信息"""
+    if not DB_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Database not available")
+    
+    db = get_db_manager()
+    relation = db.get_relation_with_details(relation_id)
+    
+    if not relation:
+        raise HTTPException(status_code=404, detail="Relation not found")
+    
+    return {
+        "relation_id": relation_id,
+        "source": relation.get("source_name"),
+        "target": relation.get("target_name"),
+        "type": relation.get("relation_type"),
+        "evidence": relation.get("evidence"),
+        "transcript": {
+            "id": relation.get("transcript_id"),
+            "filename": relation.get("transcript_filename"),
+        } if relation.get("transcript_id") else None
+    }
+
+
+@app.get("/api/v1/entities/{entity_id}/details")
+async def get_entity_details(entity_id: str):
+    """获取实体详情，包含所有提及位置"""
+    if not DB_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Database not available")
+    
+    db = get_db_manager()
+    entity = db.get_entity_with_mentions(entity_id)
+    
+    if not entity:
+        raise HTTPException(status_code=404, detail="Entity not found")
+    
+    return entity
+
+
+@app.get("/api/v1/entities/{entity_id}/evolution")
+async def get_entity_evolution(entity_id: str):
+    """分析实体的演变和态度变化"""
+    if not DB_AVAILABLE or not LLM_CLIENT_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Service not available")
+    
+    db = get_db_manager()
+    llm = get_llm_client()
+    
+    entity = db.get_entity_with_mentions(entity_id)
+    if not entity:
+        raise HTTPException(status_code=404, detail="Entity not found")
+    
+    # 分析演变
+    analysis = await llm.analyze_entity_evolution(entity["name"], entity.get("mentions", []))
+    
+    return {
+        "entity_id": entity_id,
+        "entity_name": entity["name"],
+        "mention_count": entity.get("mention_count", 0),
+        "analysis": analysis,
+        "timeline": [
+            {
+                "date": m.get("transcript_date"),
+                "snippet": m.get("text_snippet"),
+                "transcript_id": m.get("transcript_id"),
+                "filename": m.get("filename")
+            }
+            for m in entity.get("mentions", [])
+        ]
+    }
+
+
+# ==================== Phase 4: 实体管理增强 API ====================
+
+@app.get("/api/v1/projects/{project_id}/entities/search")
+async def search_entities(project_id: str, q: str):
+    """搜索实体"""
+    if not DB_AVAILABLE:
+        raise HTTPException(status_code=500, detail="Database not available")
+    
+    db = get_db_manager()
+    entities = db.search_entities(project_id, q)
+    return [{"id": e.id, "name": e.name, "type": e.type, "definition": e.definition} for e in entities]

frontend/workbench.html

@@ -584,6 +584,289 @@
         .transcript-option.active {
             background: #00d4ff22;
         }
+        
+        /* Phase 4: Agent Panel */
+        .agent-panel {
+            width: 320px;
+            background: #111;
+            border-left: 1px solid #222;
+            display: flex;
+            flex-direction: column;
+            transition: width 0.3s ease;
+        }
+        .agent-panel.collapsed {
+            width: 50px;
+        }
+        .agent-header {
+            padding: 12px 16px;
+            background: #141414;
+            border-bottom: 1px solid #222;
+            display: flex;
+            align-items: center;
+            justify-content: space-between;
+        }
+        .agent-title {
+            font-size: 0.9rem;
+            color: #00d4ff;
+            display: flex;
+            align-items: center;
+            gap: 8px;
+        }
+        .agent-toggle {
+            background: none;
+            border: none;
+            color: #666;
+            cursor: pointer;
+            font-size: 1.2rem;
+            padding: 4px;
+        }
+        .agent-toggle:hover {
+            color: #00d4ff;
+        }
+        .agent-content {
+            flex: 1;
+            display: flex;
+            flex-direction: column;
+            overflow: hidden;
+        }
+        .agent-panel.collapsed .agent-content {
+            display: none;
+        }
+        .chat-messages {
+            flex: 1;
+            overflow-y: auto;
+            padding: 16px;
+        }
+        .chat-message {
+            margin-bottom: 16px;
+            animation: fadeIn 0.3s ease;
+        }
+        @keyframes fadeIn {
+            from { opacity: 0; transform: translateY(10px); }
+            to { opacity: 1; transform: translateY(0); }
+        }
+        .chat-message.user {
+            text-align: right;
+        }
+        .chat-message.assistant {
+            text-align: left;
+        }
+        .message-bubble {
+            display: inline-block;
+            padding: 10px 14px;
+            border-radius: 12px;
+            max-width: 90%;
+            font-size: 0.9rem;
+            line-height: 1.5;
+        }
+        .chat-message.user .message-bubble {
+            background: linear-gradient(90deg, #00d4ff, #7b2cbf);
+            color: white;
+        }
+        .chat-message.assistant .message-bubble {
+            background: #1a1a1a;
+            color: #e0e0e0;
+            border: 1px solid #333;
+        }
+        .chat-input-area {
+            padding: 12px 16px;
+            border-top: 1px solid #222;
+            background: #141414;
+        }
+        .chat-input {
+            width: 100%;
+            background: #1a1a1a;
+            border: 1px solid #333;
+            border-radius: 8px;
+            padding: 10px 12px;
+            color: #e0e0e0;
+            font-size: 0.9rem;
+            resize: none;
+            outline: none;
+        }
+        .chat-input:focus {
+            border-color: #00d4ff;
+        }
+        .chat-actions {
+            display: flex;
+            gap: 8px;
+            margin-top: 8px;
+        }
+        .chat-btn {
+            background: #222;
+            border: 1px solid #333;
+            color: #888;
+            padding: 6px 12px;
+            border-radius: 6px;
+            font-size: 0.8rem;
+            cursor: pointer;
+            transition: all 0.2s;
+        }
+        .chat-btn:hover {
+            background: #333;
+            color: #00d4ff;
+            border-color: #00d4ff;
+        }
+        .chat-btn.primary {
+            background: linear-gradient(90deg, #00d4ff, #7b2cbf);
+            color: white;
+            border: none;
+        }
+        .chat-btn.primary:hover {
+            opacity: 0.9;
+        }
+        
+        /* Phase 4: Entity Card */
+        .entity-card {
+            position: fixed;
+            background: #1a1a1a;
+            border: 1px solid #333;
+            border-radius: 12px;
+            padding: 16px;
+            min-width: 280px;
+            max-width: 350px;
+            box-shadow: 0 10px 40px rgba(0,0,0,0.5);
+            z-index: 1000;
+            display: none;
+        }
+        .entity-card.show {
+            display: block;
+            animation: popIn 0.2s ease;
+        }
+        @keyframes popIn {
+            from { opacity: 0; transform: scale(0.95); }
+            to { opacity: 1; transform: scale(1); }
+        }
+        .entity-card-header {
+            display: flex;
+            align-items: center;
+            gap: 10px;
+            margin-bottom: 12px;
+            padding-bottom: 12px;
+            border-bottom: 1px solid #333;
+        }
+        .entity-card-name {
+            font-size: 1.1rem;
+            font-weight: 600;
+            color: #fff;
+        }
+        .entity-card-stats {
+            display: flex;
+            gap: 16px;
+            margin-top: 12px;
+            padding-top: 12px;
+            border-top: 1px solid #333;
+            font-size: 0.85rem;
+            color: #888;
+        }
+        .entity-card-stat {
+            display: flex;
+            align-items: center;
+            gap: 6px;
+        }
+        
+        /* Phase 4: Provenance Modal */
+        .provenance-modal {
+            position: fixed;
+            top: 0;
+            left: 0;
+            right: 0;
+            bottom: 0;
+            background: rgba(0,0,0,0.8);
+            display: none;
+            align-items: center;
+            justify-content: center;
+            z-index: 2000;
+        }
+        .provenance-modal.show {
+            display: flex;
+        }
+        .provenance-content {
+            background: #1a1a1a;
+            border: 1px solid #333;
+            border-radius: 16px;
+            width: 90%;
+            max-width: 600px;
+            max-height: 80vh;
+            overflow: hidden;
+        }
+        .provenance-header {
+            padding: 20px;
+            background: #141414;
+            border-bottom: 1px solid #333;
+            display: flex;
+            justify-content: space-between;
+            align-items: center;
+        }
+        .provenance-body {
+            padding: 20px;
+            overflow-y: auto;
+            max-height: 60vh;
+        }
+        .provenance-evidence {
+            background: #0a0a0a;
+            border-left: 3px solid #00d4ff;
+            padding: 16px;
+            margin: 16px 0;
+            border-radius: 0 8px 8px 0;
+            font-style: italic;
+            color: #ccc;
+        }
+        
+        /* Phase 4: Suggestion Card */
+        .suggestion-card {
+            background: #1a1a1a;
+            border: 1px solid #333;
+            border-radius: 8px;
+            padding: 12px;
+            margin-bottom: 12px;
+            cursor: pointer;
+            transition: all 0.2s;
+        }
+        .suggestion-card:hover {
+            border-color: #00d4ff;
+            background: #222;
+        }
+        .suggestion-type {
+            font-size: 0.75rem;
+            color: #00d4ff;
+            text-transform: uppercase;
+            margin-bottom: 4px;
+        }
+        .suggestion-title {
+            font-weight: 500;
+            margin-bottom: 4px;
+        }
+        .suggestion-desc {
+            font-size: 0.85rem;
+            color: #888;
+        }
+        
+        /* Phase 4: Low confidence entity */
+        .entity.low-confidence {
+            background: rgba(255, 193, 7, 0.3);
+            border-bottom-color: #ffc107;
+        }
+        
+        /* Typing indicator */
+        .typing-indicator {
+            display: flex;
+            gap: 4px;
+            padding: 10px 14px;
+        }
+        .typing-indicator span {
+            width: 8px;
+            height: 8px;
+            background: #666;
+            border-radius: 50%;
+            animation: typing 1.4s infinite;
+        }
+        .typing-indicator span:nth-child(2) { animation-delay: 0.2s; }
+        .typing-indicator span:nth-child(3) { animation-delay: 0.4s; }
+        @keyframes typing {
+            0%, 60%, 100% { transform: translateY(0); }
+            30% { transform: translateY(-10px); }
+        }
     </style>
 </head>
 <body>
@@ -608,6 +891,37 @@
         <div class="content-area">
             <!-- Workbench View -->
             <div id="workbenchView" class="workbench-view" style="display: flex; width: 100%;">
+                <!-- Agent Panel -->
+                <div class="agent-panel" id="agentPanel">
+                    <div class="agent-header">
+                        <div class="agent-title">
+                            <span>🤖</span>
+                            <span>Agent 助手</span>
+                        </div>
+                        <button class="agent-toggle" onclick="toggleAgentPanel()">›</button>
+                    </div>
+                    <div class="agent-content">
+                        <div class="chat-messages" id="chatMessages">
+                            <div class="chat-message assistant">
+                                <div class="message-bubble">
+                                    你好！我是 InsightFlow Agent。我可以帮你：<br><br>
+                                    • <b>问答</b>：询问项目中的任何信息<br>
+                                    • <b>合并实体</b>："把所有'客户端'合并到'App'"<br>
+                                    • <b>分析演变</b>："张总对项目的态度变化"<br>
+                                    • <b>编辑定义</b>："修改K8s的定义为..."
+                                </div>
+                            </div>
+                        </div>
+                        <div class="chat-input-area">
+                            <textarea class="chat-input" id="chatInput" rows="2" placeholder="输入指令或问题..."></textarea>
+                            <div class="chat-actions">
+                                <button class="chat-btn" onclick="loadSuggestions()">💡 建议</button>
+                                <button class="chat-btn primary" onclick="sendAgentMessage()">发送</button>
+                            </div>
+                        </div>
+                    </div>
+                </div>
+                
                 <div class="editor-panel">
                     <div class="panel-header">
                         <div style="display: flex; align-items: center; gap: 12px;">
@@ -633,7 +947,7 @@
                 <div class="graph-panel">
                     <div class="panel-header">
                         <span>🔗 知识图谱</span>
-                        <span style="font-size:0.8rem;color:#666;">右键节点编辑 | 拖拽建立关系</span>
+                        <span style="font-size:0.8rem;color:#666;">右键节点编辑 | 点击连线溯源</span>
                     </div>
                     <svg id="graph-svg"></svg>
                     <div class="entity-list" id="entityList">
@@ -703,6 +1017,38 @@
         </div>
     </div>
     
+    <!-- Entity Hover Card -->
+    <div class="entity-card" id="entityCard">
+        <div class="entity-card-header">
+            <span class="entity-type-badge" id="cardBadge">TYPE</span>
+            <span class="entity-card-name" id="cardName">Entity Name</span>
+        </div>
+        <div id="cardDefinition" style="color:#aaa;font-size:0.9rem;line-height:1.5;">暂无定义</div>
+        <div class="entity-card-stats">
+            <div class="entity-card-stat">
+                <span>📍</span>
+                <span id="cardMentions">0 次提及</span>
+            </div>
+            <div class="entity-card-stat">
+                <span>🔗</span>
+                <span id="cardRelations">0 个关系</span>
+            </div>
+        </div>
+    </div>
+    
+    <!-- Provenance Modal -->
+    <div class="provenance-modal" id="provenanceModal">
+        <div class="provenance-content">
+            <div class="provenance-header">
+                <h3>🔗 知识溯源</h3>
+                <button class="close-btn" onclick="closeProvenance()">×</button>
+            </div>
+            <div class="provenance-body" id="provenanceBody">
+                <p style="color:#666;">加载中...</p>
+            </div>
+        </div>
+    </div>
+    
     <!-- Upload Modal -->
     <div class="upload-overlay" id="uploadOverlay">
         <div class="upload-box">