# ============================================================================= # 企微IT智能服务台 — H5 员工端 AI 回复后台任务 # ============================================================================= # 背景:原 h5_send_message 在同步 HTTP 请求内 await AI 推理(Dify 3~15s), # 整条请求被阻塞,前端表现为"发送中"长时间卡顿。 # 本模块将 AI 推理移出请求,改为 asyncio 后台任务,结果经 WebSocket # 流式推回(ai_reply_chunk / ai_reply),发送瞬时完成。 # # 关键约束(详见 docs/02-需求分析/技术架构演进/员工端消息发送延时改造方案.md): # 1. 必须单 worker 运行(docker-compose --workers 1): # ws_manager 是进程内单例,多 worker 时后台任务与员工 WS 连接可能不在 # 同进程,broadcast 会静默丢失(约 50%)。 # 2. 使用独立 DB session(_get_session_factory),不可复用请求的 db # (请求返回后该 session 会被关闭)。 # ============================================================================= import asyncio import logging import os from datetime import datetime, timedelta from pathlib import Path from sqlalchemy import select from app.api.byod import _byod_keyword_prefilter from app.database import _get_session_factory from app.dependencies import get_shared_ai_handler from app.models.conversation import Conversation from app.models.message import Message from app.services.routing_service import ( routing_keyword_prefilter, detect_routing_intent, get_contact_by_category, send_contact_card, record_routing_event, _keyword_fallback_category, ) from app.services.vision_service import VisionService from app.services.ws_manager import manager as ws_manager from app.services.asset_recommend_service import get_asset_recommend_service from app.services.employee_profile_service import get_employee_profile_service from app.api.approval import APPROVAL_TEMPLATES logger = logging.getLogger(__name__) # ============================================================================= # Phase 4A: VisionService 接入 — 图片消息视觉理解 # ============================================================================= # 图片文件本地存储根目录(与 upload.py 中 UPLOAD_DIR 一致) _UPLOAD_DIR = Path(os.getenv("UPLOAD_DIR", "./uploads")) # 视觉理解置信度阈值:低于此值不注入描述(避免错误描述误导 AI) _VISION_CONFIDENCE_THRESHOLD = 0.6 def _media_url_to_local_path(media_url: str) -> Path: """将媒体 URL 路径转换为本地文件系统路径。 做什么:把 "/api/media/2026/07/13/abc.png" 转换为 "./uploads/2026/07/13/abc.png" 为什么:VisionService 需要读取原始图片字节流, 而媒体 URL 是 HTTP 访问路径,不是文件系统路径。 Args: media_url: 媒体文件 URL(如 /api/media/2026/07/13/abc.png) Returns: Path: 本地文件路径对象 """ # 去掉 URL 前缀 /api/media/,拼接到 UPLOAD_DIR # 例: "/api/media/2026/07/13/abc.png" → "2026/07/13/abc.png" relative = media_url.replace("/api/media/", "", 1) return _UPLOAD_DIR / relative async def _fetch_recent_employee_text( db, conversation_id: str, employee_id: str, within_seconds: int = 5 ) -> str: """获取最近 N 秒内员工的文字消息(Phase 4B 消息融合)。 做什么:查询同一会话中,当前图片消息之前 within_seconds 秒内, 员工发送的文本消息内容。 为什么:用户经常先打字描述问题再发截图,或先发截图再补充文字。 将文字与图片视觉描述融合后一次性传给 Dify, 避免 AI 分别处理两条消息导致上下文割裂。 Args: db: 异步 DB session conversation_id: 会话 ID employee_id: 员工企微 UserID within_seconds: 时间窗口(秒),默认 5 秒 Returns: str: 最近的员工文字消息内容(多条用换行拼接),无则返回空字符串 """ cutoff = datetime.now() - timedelta(seconds=within_seconds) stmt = ( select(Message) .where( Message.conversation_id == conversation_id, Message.sender_type == "employee", Message.sender_id == employee_id, Message.msg_type == "text", Message.created_at >= cutoff, ) .order_by(Message.created_at.desc()) .limit(3) # 最多取 3 条,避免内容过长 ) result = await db.execute(stmt) messages = result.scalars().all() if not messages: return "" # 按时间正序拼接(先发的在前) texts = [m.content for m in reversed(messages) if m.content] return "\n".join(texts) async def _enrich_image_content( db, media_url: str, original_content: str, conversation_id: str, employee_id: str, ) -> str: """用 VisionService 分析图片,生成增强后的消息内容。 做什么: 1. 从本地文件系统读取图片 2. 调用 VisionService.analyze_screenshot() 获取视觉描述 3. 查询最近 5 秒内的员工文字消息(消息融合) 4. 拼接视觉描述 + 用户文字 → 传给 Dify 为什么:Dify 文本模型无法直接"看"图片,需要先将图片转为 文字描述,再与用户输入融合后传给 Dify 推理。 降级策略: - 图片文件不存在 → 返回原始 content - VisionService 调用失败 → 返回 "我收到了您的截图,但暂时无法识别内容" - 置信度 < 0.6 → 不注入视觉描述,仅使用用户文字 Args: db: 异步 DB session media_url: 图片 URL(如 /api/media/2026/07/13/abc.png) original_content: 原始消息内容(如 "[图片] 截图") conversation_id: 会话 ID employee_id: 员工企微 UserID Returns: str: 增强后的消息内容(视觉描述 + 用户文字) """ # 1. 读取本地图片文件 local_path = _media_url_to_local_path(media_url) if not local_path.exists(): logger.warning(f"图片文件不存在: {local_path} (media_url={media_url})") return original_content try: image_bytes = local_path.read_bytes() except Exception as e: logger.error(f"读取图片文件失败: {local_path} - {e}") return original_content # 2. 调用 VisionService 分析截图 vision_service = VisionService() try: result = await vision_service.analyze_screenshot( image_bytes, conversation_id ) description = result.get("description", "") confidence = result.get("confidence", 0.0) logger.info( f"VisionService 分析完成: conversation={conversation_id}, " f"confidence={confidence:.2f}, desc_len={len(description)}" ) # 3. 注入视觉描述到会话上下文(供后续多轮对话使用) if description and confidence >= _VISION_CONFIDENCE_THRESHOLD: await vision_service.inject_to_conversation_context( description, conversation_id ) except Exception as e: logger.error(f"VisionService 调用异常: {e}") description = "" confidence = 0.0 finally: await vision_service.close() # 4. 消息融合:查询最近 5 秒内员工的文字消息 recent_text = await _fetch_recent_employee_text( db, conversation_id, employee_id, within_seconds=5 ) # 5. 拼接增强内容 # 格式:[视觉描述] + [用户最近文字] + [原始消息内容] parts = [] if description and confidence >= _VISION_CONFIDENCE_THRESHOLD: parts.append(f"[用户发送了截图,视觉理解结果] {description}") if recent_text: parts.append(f"[用户最近的文字描述] {recent_text}") # 原始内容如果不是纯占位符(如"[图片] 截图"),也加入 if original_content and not original_content.startswith("[图片]"): parts.append(original_content) if not parts: # 降级:视觉分析失败且无文字补充 return "我收到了您的截图,但暂时无法识别内容,请描述一下您遇到的问题。" return "\n".join(parts) async def _persist_and_push_solution( db, conversation, employee_id: str, solution, ): """处理图谱命中的解决方案(图谱查询结果) 做什么: 1. 创建AI回复消息记录(图谱命中的解决方案) 2. 通过WS推送给员工端 为什么:图谱命中的解决方案直接返回,不需要调用Dify Args: db: 数据库会话 conversation: 会话对象 employee_id: 员工ID solution: SolutionResult图谱查询结果 """ from app.models.message import Message from app.services.ws_manager import manager as ws_manager # 1. 创建AI回复消息记录 message = Message( conversation_id=conversation.id, sender_type="ai", sender_id="graph", content=solution.solution, msg_type="text", ) db.add(message) # 更新会话计数 conversation.ai_substantive_reply_count += 1 conversation.updated_at = datetime.now() await db.commit() # 2. 构建推送数据 msg_data = { "id": str(message.id), "conversation_id": str(conversation.id), "sender_type": "ai", "sender_id": "graph", "sender_name": "智能助手", "content": solution.solution, "msg_type": "text", "created_at": message.created_at.isoformat(), "reply_source": "graph_hit", # 标记来源为图谱命中 } # 3. 通过WS推送给员工端 try: await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_reply", "data": msg_data, }) logger.info( f"图谱命中推送成功: employee={employee_id}, " f"solution={solution.action_name}" ) except Exception as push_err: logger.error(f"图谱命中推送失败: {push_err}") async def _persist_and_push( db, conversation: Conversation, employee_id: str, content: str, is_guidance: bool, should_count: bool, should_transfer: bool, dify_conversation_id, ): """持久化 AI 回复并推送给员工端 + 广播坐席端。 做什么: 1. 存 AI 消息到 DB 2. 更新会话状态(dify 上下文 / 计数 / 转人工) 3. 经 WS 向员工推 ai_reply 终态(前端据此替换打字机气泡) 4. 经 WS 向坐席端广播 new_message + conversation_updated 为什么:把"落库 + 推送"封装为单点,供同步路径与流式路径复用。 """ # 1. 存 AI 消息 ai_message = Message( conversation_id=conversation.id, sender_type="ai", sender_id="ai_bot", sender_name="Duckula(达寇拉)", content=content, msg_type="text", is_read=True, ) db.add(ai_message) await db.flush() # 2. 更新会话状态 if dify_conversation_id: conversation.dify_conversation_id = dify_conversation_id if should_count: conversation.ai_substantive_reply_count += 1 if should_transfer: conversation.status = "queued" conversation.updated_at = datetime.now() db.add(conversation) await db.flush() await db.commit() # 3. 推 ai_reply 终态给员工(前端替换打字机气泡) await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_reply", "data": { "message_id": str(ai_message.id), "conversation_id": str(conversation.id), "sender_type": "ai", "sender_id": "ai_bot", "sender_name": "Duckula(达寇拉)", "content": content, "msg_type": "text", "is_guidance": is_guidance, "ai_reply_count": conversation.ai_substantive_reply_count, "can_call_agent": conversation.ai_substantive_reply_count >= 3, "conversation_status": conversation.status, }, }) # 4. 广播坐席端(new_message + conversation_updated) try: await ws_manager.broadcast({ "type": "new_message", "data": { "conversation_id": str(conversation.id), "message_id": str(ai_message.id), "sender_type": "ai", "sender_id": "ai_bot", "sender_name": "Duckula(达寇拉)", "content": content, "msg_type": "text", }, }) await ws_manager.broadcast({ "type": "conversation_updated", "data": { "conversation_id": str(conversation.id), "status": conversation.status, "assigned_agent_id": str(conversation.assigned_agent_id) if conversation.assigned_agent_id else None, }, }) except Exception as ws_err: # WS 广播失败不阻塞消息存储,只记录 warning logger.warning(f"WS 广播 AI 回复给坐席失败(消息已存储): {ws_err}") async def _persist_and_push_structured( db, conversation: Conversation, employee_id: str, result: dict, ): """持久化结构化 AI 回复并推送给员工端 + 广播坐席端(v2.0 双 WS 通道)。 改造后的核心变化(2026-07-13): - Dify 返回 JSON {text, action, options},后端解析后同时发两条 WS: ① ai_reply → 聊天气泡(text + options) ② dynamic_recommend → 侧边栏推荐(action 卡片) - 两条消息同一时刻发出,零时间差到达 - 文字明确引用侧边栏内容(如"右侧已为您准备好入口"),语义强关联 命中判断规则: - 结构化回复且有 action 或 options → 视为命中(AI 在主动引导) - 纯文本回复 → 走原有 _check_knowledge_hit 判断 Args: db: 异步 DB session conversation: 当前会话对象 employee_id: 员工企微 UserID result: get_structured_reply() 返回的结构化结果 """ text = result.get("text", "") # ★ 防御性类型保护:确保 content 始终是 String # 如果 Dify 返回的 text 是 dict/list,WS 推送后前端会显示 [object Object] if not isinstance(text, str): import json as _json text = _json.dumps(text, ensure_ascii=False) if text else "" logger.warning(f"_persist_and_push_structured: text 非 String 类型,已转换: {text[:80]}...") action = result.get("action") options = result.get("options") hit = result.get("hit", False) is_structured = result.get("is_structured", False) dify_conv_id = result.get("conversation_id") # Phase 6A: 提取诊断阶段 diagnosis_stage = result.get("diagnosis_stage") # 结构化回复且有 action 或 options → 视为命中(AI 在主动引导/推荐) if is_structured and (action or options): hit = True # Phase 6A: 基于 diagnosis_stage 调整会话状态 # escalating → AI 建议转人工 # resolved → AI 认为问题已解决 if diagnosis_stage == "escalating": hit = False # 不计为有效回复,触发转人工 elif diagnosis_stage == "resolved": hit = True # 计为有效回复 should_count = hit should_transfer = not hit # 确定消息类型 # v2.4 修复:有 action(审批卡片)时,强制设置为 ai_structured # 确保前端能渲染审批卡片入口,不依赖 Dify 返回的 is_structured 字段 if action or options or is_structured: msg_type = "ai_structured" else: msg_type = "text" # v2.5 调试日志 logger.info(f"[DEBUG] msg_type = {msg_type}, action = {bool(action)}, options = {bool(options)}, is_structured = {is_structured}") # 构建 extra_data(存储 options 和 action 供前端渲染) extra_data = {} if options: extra_data["options"] = options # 为审批卡片注入标准化 card_data(替换原有的分散匹配逻辑) if action: # v3.2 修复:降级路径已构建好 card_data 时,跳过重复匹配(避免 approval_type=None 误报"匹配失败") if action.get("card_data"): matched_card = action["card_data"] options = matched_card.get("options", []) if options and not action.get("url"): action["url"] = options[0].get("url", "") logger.info(f"[ApprovalMatcher] 使用预构建 card_data: {matched_card.get('title')}") else: approval_type = action.get("approval_type") title = action.get("title") # v3.0 重构:委托 ApprovalMatcher 统一完成模板匹配 + 卡片构建 from app.services.approval_matcher import get_approval_matcher matcher = get_approval_matcher() matched_card = matcher.match_and_build_card(approval_type, title) if matched_card: # 注入 URL 供 approve-direct-card 兼容路径使用 options = matched_card.get("options", []) if options: action["url"] = options[0].get("url", "") action["card_data"] = matched_card logger.info(f"[ApprovalMatcher] 匹配成功: {approval_type} -> card_type={matched_card.get('card_type')}") else: # v4.0 P1-4 后:matcher 仅在 approval_type 为空时返回 None(其余情况兜底全量卡片) logger.info(f"[ApprovalMatcher] approval_type 为空,无卡片: title={title}") extra_data["action"] = action # ★ 调试日志:打印推送到前端的 extra_data 内容 logger.info(f"[DEBUG] 推送到前端的 extra_data: {extra_data}") if extra_data.get("action"): logger.info(f"[DEBUG] extra_data.action.approval_type = {extra_data['action'].get('approval_type')}") # 1. 存 AI 消息 ai_message = Message( conversation_id=conversation.id, sender_type="ai", sender_id="ai_bot", sender_name="Duckula(达寇拉)", content=text, msg_type=msg_type, extra_data=extra_data if extra_data else None, is_read=True, ) db.add(ai_message) await db.flush() # 2. 更新会话状态 if dify_conv_id: conversation.dify_conversation_id = dify_conv_id if should_count: conversation.ai_substantive_reply_count += 1 if should_transfer: conversation.status = "queued" # Phase 6A: 将 diagnosis_stage 存入 tags(无需迁移,利用现有 JSON 字段) if diagnosis_stage: tags = conversation.tags or {} tags["diagnosis_stage"] = diagnosis_stage tags["diagnosis_updated_at"] = datetime.now().isoformat() conversation.tags = tags conversation.updated_at = datetime.now() db.add(conversation) await db.flush() await db.commit() # 3. 推 ai_reply 给员工端(聊天气泡:text + options) # 添加异常处理,避免整体失败 try: await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_reply", "data": { "message_id": str(ai_message.id), "conversation_id": str(conversation.id), "sender_type": "ai", "sender_id": "ai_bot", "sender_name": "Duckula(达寇拉)", "content": text, "msg_type": msg_type, "extra_data": extra_data if extra_data else None, "is_guidance": False, "ai_reply_count": conversation.ai_substantive_reply_count, "can_call_agent": conversation.ai_substantive_reply_count >= 3, "conversation_status": conversation.status, # Phase 6A: 诊断阶段(前端可据此调整 UI/提示) "diagnosis_stage": diagnosis_stage, }, }) except Exception as emp_err: # 员工端推送失败不应该导致整个任务失败 logger.warning(f"员工端 AI 回复推送失败(不影响坐席端): {emp_err}") # 4. 推 dynamic_recommend 给员工端侧边栏(仅当 action 非空且不是审批类型时) # 与 ai_reply 同一时刻发出 → 零时间差到达 # v2.3 修改:审批类型只推送到消息气泡(左边),不推送到侧边栏(右边) if action and not action.get("approval_type"): # 为审批卡片注入运维平台跳转URL(实现免登录跳转) approval_type = action.get("approval_type") action_url = "" location = "运维平台" # 优先精确匹配:直接用 approval_type 查找模板 if approval_type and approval_type in APPROVAL_TEMPLATES: template = APPROVAL_TEMPLATES[approval_type] action_url = template.get("url", "") location = template.get("location", "运维平台") # 关键字匹配:当精确匹配失败时,通过关键字查找模板 # Dify返回的 approval_type 可能是中文分类名(如"账号权限申请"、"VPN账号申请") elif approval_type: for template_id, template in APPROVAL_TEMPLATES.items(): keywords = template.get("keywords", []) # 检查 approval_type 是否包含任意一个关键字 if any(kw.lower() in approval_type.lower() for kw in keywords): action_url = template.get("url", "") location = template.get("location", "运维平台") break # v2.2 新增:根据 Dify 返回的 approval_type 设置 filtered_options filtered_options = [] if approval_type: filtered_options = [approval_type] recommend_data = { "recommend_id": f"rec_{ai_message.id}", "card_type": action.get("type", "approval_card"), "title": action.get("title", ""), "description": action.get("description", ""), "approval_type": action.get("approval_type"), "filtered_options": filtered_options, # v2.2: 精确匹配的选项 "action_url": action_url, # 运维平台跳转URL "action_label": f"打开{location}" if action_url else "打开审批表单", # 按钮文字 "location": location, # 平台名称 "confidence": action.get("confidence", 0.85), "message_id": str(ai_message.id), "conversation_id": str(conversation.id), } try: await ws_manager.broadcast_to_employees([employee_id], { "type": "dynamic_recommend", "data": recommend_data, }) logger.info( f"动态推荐已推送: employee={employee_id}, " f"card_type={recommend_data['card_type']}, " f"title={recommend_data['title']}" ) except Exception as rec_err: logger.warning(f"员工端动态推荐推送失败: {rec_err}") # 5. 广播坐席端(new_message + conversation_updated) try: await ws_manager.broadcast({ "type": "new_message", "data": { "conversation_id": str(conversation.id), "message_id": str(ai_message.id), "sender_type": "ai", "sender_id": "ai_bot", "sender_name": "Duckula(达寇拉)", "content": text, "msg_type": msg_type, "extra_data": extra_data if extra_data else None, }, }) await ws_manager.broadcast({ "type": "conversation_updated", "data": { "conversation_id": str(conversation.id), "status": conversation.status, "assigned_agent_id": ( str(conversation.assigned_agent_id) if conversation.assigned_agent_id else None ), }, }) except Exception as ws_err: logger.warning(f"WS 广播结构化 AI 回复给坐席失败(消息已存储): {ws_err}") async def _handle_byod_query(db, conversation, employee_id, content): """处理 BYOD 自备电脑补贴查询。 在 H5 聊天消息流中拦截 BYOD 关键词后执行资格检查,并以 byod_card 卡片消息形式推送给员工端(前端 MessageBubble 据 msg_type 渲染 ByodSubsidyCard)。 流程: 1. 通过 WecomService 获取员工岗位(position) 2. 与 BYOD 资格清单匹配(_match_position) 3. 创建 byod_card 类型 AI 消息并落库 4. 经 WS 推送 ai_reply 给员工端(携带 extra_data.byod_result) 5. 广播 new_message + conversation_updated 给坐席端(与 _persist_and_push 一致) Args: db: 异步 DB session(process_h5_ai_reply 的 factory session) conversation: 当前会话对象(Conversation) employee_id: 员工企微 UserID content: 用户消息原文(用于日志) """ # 延迟导入避免循环依赖(byod 模块注册路由时可能引用 app.main) from app.api.byod import _match_position, BYOD_APPLICATION_URL, BYOD_NOTES, BYOD_REGISTER_NOTES from app.services.wecom_service import WecomService # 1. 获取员工岗位(企微通讯录 API) position = "" try: wecom_service = WecomService() try: user_info = await wecom_service.get_user_info(employee_id) position = user_info.get("position", "") finally: await wecom_service.close() except Exception as e: logger.error(f"BYOD: 获取员工岗位失败: {e}") # 2. 岗位匹配(返回: 是否匹配, 匹配岗位, 匹配类别) eligible, matched_pos, matched_category = _match_position(position) # 3. 构建 BYOD 结果数据 # 字段与前端 ByodSubsidyCard.vue props 完全一致: # eligible / position / matched_category / application_url / notes / reason byod_result = { "eligible": eligible, "has_subsidy": eligible, "position": position, "matched_category": matched_category, "application_url": BYOD_APPLICATION_URL, # 所有岗位都提供链接 "notes": BYOD_NOTES if eligible else BYOD_REGISTER_NOTES, "reason": ( "" if eligible else f"您的岗位「{position}」不在自备电脑补贴资格清单中,可进行自备电脑登记(无补贴)" ), } # 4. 展示文本(AI 气泡的 content,卡片下方不直接展示,但会话列表/坐席端可见) if eligible: display_text = f"您岗位为「{position}」,符合自备电脑补贴申请资格" else: display_text = f"您岗位为「{position}」,可进行自备电脑登记(无补贴)" # 5. 创建 AI 消息(byod_card 类型,携带 byod_result) ai_message = Message( conversation_id=conversation.id, sender_type="ai", sender_id="ai_bot", sender_name="Duckula(达寇拉)", content=display_text, msg_type="byod_card", extra_data={"byod_result": byod_result}, is_read=True, ) db.add(ai_message) await db.flush() # 6. 更新会话状态(计数 + 时间,BYOD 视为一次实质性 AI 回复) conversation.ai_substantive_reply_count += 1 conversation.updated_at = datetime.now() db.add(conversation) await db.flush() await db.commit() # 7. 推送 ai_reply 给员工端(前端据 msg_type="byod_card" 渲染卡片) await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_reply", "data": { "message_id": str(ai_message.id), "conversation_id": str(conversation.id), "sender_type": "ai", "sender_id": "ai_bot", "sender_name": "Duckula(达寇拉)", "content": display_text, "msg_type": "byod_card", "extra_data": {"byod_result": byod_result}, "is_guidance": False, "ai_reply_count": conversation.ai_substantive_reply_count, "can_call_agent": conversation.ai_substantive_reply_count >= 3, "conversation_status": conversation.status, }, }) # 8. 广播坐席端(new_message + conversation_updated,与 _persist_and_push 一致) try: await ws_manager.broadcast({ "type": "new_message", "data": { "conversation_id": str(conversation.id), "message_id": str(ai_message.id), "sender_type": "ai", "sender_id": "ai_bot", "sender_name": "Duckula(达寇拉)", "content": display_text, "msg_type": "byod_card", "extra_data": {"byod_result": byod_result}, }, }) await ws_manager.broadcast({ "type": "conversation_updated", "data": { "conversation_id": str(conversation.id), "status": conversation.status, "assigned_agent_id": ( str(conversation.assigned_agent_id) if conversation.assigned_agent_id else None ), }, }) except Exception as ws_err: logger.warning(f"BYOD: WS 广播给坐席失败: {ws_err}") logger.info( f"BYOD 查询完成: employee_id={employee_id}, position={position}, " f"eligible={eligible}, matched_category={matched_category}" ) async def _handle_routing( db, conversation: Conversation, employee_id: str, content: str, ) -> bool: """处理非IT业务路由推荐。 在 H5 聊天消息流中拦截路由关键词后调用 Dify 统一意图识别, 判定为 non_it_routing 且 routing_confidence ≥ 阈值时发送名片三段式消息。 流程: 1. 调用 Dify 统一意图识别(detect_routing_intent) 2. 检查 intent_type == "non_it_routing" && routing_confidence ≥ 阈值 → YES: 查询联系人 → 发送名片三段式消息 → 记录路由事件 → 返回 True → NO: 返回 False(继续走正常 AI 流程) 3. Dify 调用失败 → 关键词降级兜底(按 ROUTING_KEYWORD_TO_CATEGORY 映射) Args: db: 异步 DB session(process_h5_ai_reply 的 factory session) conversation: 当前会话对象(Conversation) employee_id: 员工企微 UserID content: 用户消息原文 Returns: bool: True 表示已发送路由名片(应 return 中断后续流程), False 表示未触发路由(继续走正常 AI 流程) """ from app.config import settings # 1. 调用 Dify 统一意图识别 try: result = await detect_routing_intent(content, employee_id) intent_type = result.get("intent_type", "chitchat") business_category = result.get("business_category") routing_confidence = result.get("routing_confidence", 0.0) # 如果是审批意图,不拦截(让审批流程处理) if intent_type == "approval": return False logger.info( f"路由意图检测(Dify): intent_type={intent_type}, " f"business_category={business_category}, " f"routing_confidence={routing_confidence}" ) # 2. 检查是否触发路由推荐 threshold = settings.routing_confidence_threshold if intent_type != "non_it_routing" or routing_confidence < threshold: # 置信度不足或非路由意图,走正常 AI 流程 return False if not business_category: logger.warning("路由意图为 non_it_routing 但 business_category 为空,跳过") return False except Exception as e: logger.warning(f"Dify 路由意图识别失败,降级为关键词匹配: {e}") # 3. 降级为关键词匹配 business_category = _keyword_fallback_category(content) if not business_category: # 关键词也未命中,走正常 AI 流程 return False routing_confidence = 0.75 # 降级兜底给一个略高于阈值的置信度 logger.info(f"路由意图检测(兜底): business_category={business_category}") # 4. 查询联系人 contact = await get_contact_by_category(db, business_category) if not contact: logger.warning(f"未找到 {business_category} 类别的联系人,跳过路由推荐") return False # 5. 构建路由说明文本 category_display = business_category.replace("行政-物业", "物业") reason = ( f"您的问题属于{category_display}业务范畴,不在IT服务台服务范围内 😊\n\n" f"为您推荐{category_display}服务相关联系人,您可以直接点击名片联系TA:" ) # 6. 发送名片三段式消息 await send_contact_card( db=db, conversation=conversation, employee_id=employee_id, contact=contact, reason=reason, business_category=business_category, routing_confidence=routing_confidence, ) # 7. 记录路由事件(P1) await record_routing_event( db=db, conversation_id=str(conversation.id), employee_id=employee_id, message_content=content, business_category=business_category, routing_confidence=routing_confidence, contact=contact, ) return True async def _enrich_with_last_ai_context(db, conversation_id: str, content: str) -> str: """为简短回复拼接对话上下文,弥补 Dify 工作流缺少「对话历史」节点。 v2.3 改进(相对于 v2.2): - 移除 15 字符硬限制 → 50 字符宽松阈值(问句/换行/长消息自动跳过) - 查询最近 10 条消息(用户+AI)→ 构建完整对话摘要,含用户原始问题 - 不再依赖 extra_data.options 判断,对所有简短回复尝试拼接 - 跳过刚保存的当前消息避免重复(当前内容已作为 query 单独传给 Dify) 触发条件:消息不含问号/换行、长度 <= 50 字符 → 可能是选项选择/简短回答。 后续:Dify 工作流配置对话历史节点后,可将 `MAX_CONTEXT_LENGTH` 设为 0 来禁用此修复。 返回:拼接后的消息(如果不需要拼接则返回原内容) """ # 宽松的启发式判断:不含问号、不含换行、<= 50 字符 → 可能是简短回答 if "?" in content or "?" in content or "\n" in content or len(content) > 50: return content try: # 1. 查询最近 10 条消息(按时间倒序索取最新),含用户和 AI stmt = ( select(Message.content, Message.sender_type, Message.created_at) .where(Message.conversation_id == conversation_id) .order_by(Message.created_at.desc()) .limit(10) ) result = await db.execute(stmt) rows = list(result.all()) if not rows or len(rows) < 2: return content # 消息太少,无法构建有意义上下文 # 2. 反转顺序:最早 → 最新 rows.reverse() # 3. 跳过最后一条员工消息 → 即刚刚保存的当前消息(避免在上下文中重复) # content 已作为 query 单独发给 Dify,不应出现在上下文中 if rows and rows[-1][1] == "employee": rows = rows[:-1] if len(rows) < 2: return content # 去掉当前消息后没剩几条,不拼接 # 4. 构建对话摘要(最多保留最近 8 条,避免 prompt 过长) context_lines = [] for row_text, row_sender, _ in rows[-8:]: if not row_text: continue role = "用户" if row_sender == "employee" else "AI助手" context_lines.append(f"{role}: {row_text}") if not context_lines: return content context = "\n".join(context_lines) return ( f"【对话上下文】\n{context}\n\n" f"请根据以上对话历史回答用户的以下消息:{content}" ) except Exception as e: logger.warning(f"上下文拼接失败,使用原消息: {e}") return content async def _push_asset_recommends( db, employee_id: str, message: str, dify_result: dict, ): """v3.0 资产推荐推送 - 独立于对话的运维触达通道 功能: 1. L1: 从关键词匹配资产(与当前问题相关) 2. L2: 从画像触发运维提醒(与问题无关) 3. L3: 角色通用资源推荐 Args: db: 数据库会话 employee_id: 员工 ID message: 用户消息(用于关键词匹配) dify_result: Dify 返回结果(包含 intent 等信息) """ try: asset_service = get_asset_recommend_service() profile_service = get_employee_profile_service() # 1. L1: 关键词匹配(从用户消息中提取关键词) l1_recs = asset_service.match_keywords(message) # 2. L2+L3: 画像匹配(需要获取员工画像) # 为避免每次都调用第三方 API,先尝试获取画像 # 画像获取失败时只推送 L1 profile = None try: profile = await profile_service.get_profile(employee_id) profile_dict = { 'huorong_version': profile.huorong_version, 'huorong_virusdb_date': profile.huorong_virusdb_date, 'huorong_offline_days': profile.huorong_offline_days, 'unionsoft_patches_missing': profile.unionsoft_patches_missing, 'unionsoft_violations': profile.unionsoft_violations, } l2_recs = asset_service.match_profile_triggers(profile_dict) # L3: 角色通用推荐 role = profile.position or '' l3_recs = asset_service.get_by_role(role) for rec in l3_recs: rec.layer = 'L3' rec.layer_label = '常用资源' rec.relevance = 'low' except Exception as e: logger.warning(f"[AssetRecommend] 获取画像失败: {e}") l2_recs = [] l3_recs = [] # 3. 合并所有推荐(去重) all_recs = l1_recs + l2_recs + l3_recs if not all_recs: logger.debug(f"[AssetRecommend] 无推荐: employee={employee_id}") return # 4. 构建 WS 消息并推送(添加异常处理避免影响主流程) try: ws_msg = asset_service.build_ws_message(all_recs) await ws_manager.broadcast_to_employees([employee_id], ws_msg) logger.info( f"[AssetRecommend] 已推送: employee={employee_id}, " f"L1={len(l1_recs)}, L2={len(l2_recs)}, L3={len(l3_recs)}" ) except Exception as ws_err: logger.warning(f"[AssetRecommend] WS推送失败(不影响主流程): {ws_err}") except Exception as e: logger.error(f"[AssetRecommend] 推送失败: {e}", exc_info=True) # 资产推荐失败不影响主对话流程 # ============================================================================= # 管线步骤函数(v4.0 批次 3:process_h5_ai_reply 管线化重构) # ============================================================================= # 设计:主函数从 200+ 行/11 对 try/except 收敛为 ~40 行编排代码, # 每步一个函数,步骤内部自管异常,主函数零嵌套。 # 行为承诺:与原 v3.2 实现外部行为一致(仅结构调整 + v3.0 降级结果补 type 字段)。 # ============================================================================= async def _step_load_conversation(db, conversation_id: str): """步骤1:加载会话(重试 3 次,处理事务未提交竞态)。""" for attempt in range(3): conversation = await db.get(Conversation, conversation_id) if conversation: return conversation if attempt < 2: await asyncio.sleep(0.5) # 刷新 session 以看到已提交的数据 await db.rollback() logger.warning(f"后台 AI 任务:会话不存在(重试3次后) {conversation_id}") return None async def _step_byod_intercept(db, conversation, employee_id, content, msg_type) -> bool: """步骤2:BYOD 关键词拦截。命中返回 True(终止管线)。""" if msg_type == "text" and _byod_keyword_prefilter(content): await _handle_byod_query(db, conversation, employee_id, content) return True return False async def _step_routing_intercept(db, conversation, employee_id, content, msg_type) -> bool: """步骤3:非IT业务路由拦截。命中并发送名片返回 True(终止管线)。""" if msg_type != "text" or not routing_keyword_prefilter(content): return False return await _handle_routing(db, conversation, employee_id, content) async def _step_local_quick_reply(db, conversation, employee_id, content, msg_type, dify_conversation_id) -> bool: """步骤4:本地快判断(打招呼/呼叫人工)。命中返回 True(终止管线)。""" if msg_type != "text": return False ai_handler = get_shared_ai_handler() if not (ai_handler.is_greeting(content) or ai_handler.is_call_human(content)): return False result = await ai_handler.handle_message( content=content, dify_conversation_id=dify_conversation_id, user_id=employee_id, ) await _persist_and_push( db, conversation, employee_id, result.content, result.is_guidance, result.should_count, result.should_transfer, result.dify_conversation_id, ) return True async def _step_enrich_image(db, content, msg_type, media_url, conversation_id, employee_id) -> str: """步骤5:图片消息增强(VisionService 分析 -> 视觉描述融合)。失败降级为原文。""" if msg_type != "image" or not media_url: return content logger.info(f"图片消息检测: conversation={conversation_id}, media_url={media_url}") try: enriched = await _enrich_image_content( db=db, media_url=media_url, original_content=content, conversation_id=conversation_id, employee_id=employee_id, ) logger.info( f"图片内容增强完成: original_len={len(content)}, enriched_len={len(enriched)}" ) return enriched except Exception as vision_err: logger.error(f"VisionService 处理失败,降级为纯文本: {vision_err}") # 降级:使用原始 content,AI 会收到 "[图片] 截图" 这样的占位符 return content async def _step_graph_shortcut(db, conversation, employee_id, enriched_content, msg_type) -> bool: """步骤6:Neo4j 图谱短路。命中返回 True(终止管线)。异常降级继续。""" if msg_type != "text" or not enriched_content: return False try: from app.services.graph_query_service import get_graph_query_service from app.services.neo4j_client import get_neo4j_client neo4j_client = await get_neo4j_client() if not neo4j_client: return False graph_service = await get_graph_query_service(neo4j_client) solution = ( await graph_service.find_solution_by_question(enriched_content) if graph_service else None ) if not solution: return False logger.info( f"图谱命中: question={enriched_content[:30]}, solution={solution.action_name}" ) # 直接返回图谱解决方案,跳过 Dify 调用 await _persist_and_push_solution(db, conversation, employee_id, solution) return True except Exception as graph_err: # 图谱查询失败不阻断,继续原有 Dify 流程 import traceback logger.warning(f"图谱查询异常(降级继续): {graph_err}\n{traceback.format_exc()}") return False def _build_keyword_fallback_result(matched_card, conversation, result=None, source="keyword_fallback"): """构建关键词降级结果(v3.0 超时降级 / v3.1 无action降级共用)。 v4.0 批次 3 合并:统一补 "type": "approval_card"(v3.0 原缺此字段, 前端 approve-direct-card 渲染依赖它,属于缺陷修复)。 """ options = matched_card.get("options") or [] return { "text": f"我来帮您提交{matched_card.get('title', '审批')},请点击下方卡片。", "action": { "type": "approval_card", "card_data": matched_card, "url": options[0].get("url", "") if options else "", }, "options": None, "hit": True, "conversation_id": (result or {}).get("conversation_id") or conversation.dify_conversation_id, "is_structured": True, "diagnosis_stage": "recommending" if result else None, "response_time_ms": (result or {}).get("response_time_ms", 0), "source": source, } async def _step_notify_failure(conversation_id: str, employee_id: str, message: str): """步骤:统一失败通知(ai_reply_failed WS 推送)。""" try: await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_reply_failed", "data": {"conversation_id": conversation_id, "message": message}, }) except Exception: # 推送失败也无所谓,员工端 3 秒轮询兜底 pass async def _step_call_dify(db, conversation, employee_id, conversation_id, content, enriched_content, dify_conversation_id): """步骤7:Dify 主推理(thinking 指示器 + 30s 超时 + 关键词降级)。 返回结构化结果 dict; 超时且降级失败时内部推送 ai_reply_failed 并返回 None(终止管线); 超时但降级成功时已推送降级卡片,返回 None(终止管线)。 """ # 1. 推送 "正在思考..." 指示器(员工气泡动画 + 坐席状态指示) try: await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_thinking", "data": {"conversation_id": conversation_id}, }) except Exception as thinking_err: logger.warning(f"推送 AI 思考指示器失败: {thinking_err}") try: await ws_manager.broadcast({ "type": "ai_thinking", "data": {"conversation_id": conversation_id, "employee_id": employee_id}, }) except Exception: pass # 坐席端通知失败不影响主流程 # 2. 启动延迟 "仍在思考" 后台任务(15 秒后触发,Dify 返回则取消) async def _push_still_thinking(): await asyncio.sleep(15) await ws_manager.broadcast_to_employees([employee_id], { "type": "ai_thinking", "data": {"conversation_id": conversation_id, "status": "still_thinking"}, }) thinking_task = asyncio.create_task(_push_still_thinking()) # 3. 调用 Dify(blocking + 30s 硬超时) ai_handler = get_shared_ai_handler() try: result = await asyncio.wait_for( ai_handler.ai_service.get_structured_reply( message=enriched_content, conversation_id=dify_conversation_id or conversation.dify_conversation_id, user_id=employee_id, ), timeout=30, ) except asyncio.TimeoutError: # v3.0: Dify 超时 -> 关键词降级匹配 thinking_task.cancel() logger.warning(f"Dify 30 秒超时: conversation={conversation_id},尝试关键词降级") from app.services.approval_matcher import get_approval_matcher matched_card = get_approval_matcher().match_by_keywords(content) if matched_card: logger.info(f"[Fallback] 关键词降级成功: {matched_card.get('title')}") fallback_result = _build_keyword_fallback_result(matched_card, conversation) try: await _persist_and_push_structured(db, conversation, employee_id, fallback_result) return None # 已推送降级卡片,终止管线 except Exception as fallback_err: logger.error(f"[Fallback] 降级推送失败: {fallback_err}") # 降级也失败 -> 建议转人工 await _step_notify_failure(conversation_id, employee_id, "AI 响应时间较长,建议转人工坐席处理。") return None # 4. 取消 "仍在思考" 任务(Dify 已返回) thinking_task.cancel() try: await thinking_task # 等待 task 真正取消,避免 warning except asyncio.CancelledError: pass # 5. v3.1: Dify 返回但无 action(如诊断 escalate)-> 关键词降级 if not result.get("action"): from app.services.approval_matcher import get_approval_matcher matched_card = get_approval_matcher().match_by_keywords(content) if matched_card: logger.info(f"[Fallback-v3.1] Dify 无action但关键词命中: {matched_card.get('title')}") result = _build_keyword_fallback_result(matched_card, conversation, result, "keyword_fallback_v3") return result async def _step_persist(db, conversation, employee_id, result): """步骤8:持久化 + 双 WS 推送(ai_reply + dynamic_recommend)。异常不中断管线。""" try: await _persist_and_push_structured(db, conversation, employee_id, result) except Exception as persist_err: logger.error(f"[Persist] AI回复持久化失败: {persist_err}", exc_info=True) async def _step_assets(db, employee_id, content, result): """步骤9:资产推荐推送(L1/L2/L3 分层,独立运维触达通道)。异常不中断管线。""" try: await _push_asset_recommends(db, employee_id, content, result) except Exception as asset_err: logger.error(f"[Asset] 资产推荐推送失败: {asset_err}", exc_info=True) # ============================================================================= # 管线编排主函数(v4.0 批次 3 重构版) # ============================================================================= async def process_h5_ai_reply( conversation_id: str, employee_id: str, content: str, dify_conversation_id=None, msg_type: str = "text", media_url: str = None, ): """H5 发送消息后的 AI 回复处理(asyncio.create_task 入口)。 v4.0 批次 3 管线化重构:原 12 步/11 对 try/except 编排为 9 个步骤函数, 主函数仅最外层 1 个 try/except,行为与 v3.2 外部表现一致。 流程: 1. _step_load_conversation 加载会话(重试3次) 2. _step_byod_intercept BYOD 关键词拦截 3. _step_routing_intercept 非IT业务路由拦截(名片推荐) 4. _step_local_quick_reply 本地快判断(打招呼/呼叫人工) 5. _step_enrich_image 图片消息 VisionService 增强 5b. _enrich_with_last_ai_context 简短回复上下文拼接(v2.3) 6. _step_graph_shortcut Neo4j 图谱短路 7. _step_call_dify Dify 主推理(含超时/无action关键词降级) 8. _step_persist 持久化 + 双 WS 推送 9. _step_assets 资产推荐推送 Args: conversation_id: 会话 ID employee_id: 员工企微 UserID content: 消息文本内容 dify_conversation_id: Dify 会话 ID(用于多轮上下文) msg_type: 消息类型(text/image/file),默认 text media_url: 媒体文件 URL(图片消息时使用) """ factory = _get_session_factory() async with factory() as db: try: conversation = await _step_load_conversation(db, conversation_id) if not conversation: return # 前置拦截管线(任一命中即终止) if await _step_byod_intercept(db, conversation, employee_id, content, msg_type): return if await _step_routing_intercept(db, conversation, employee_id, content, msg_type): return if await _step_local_quick_reply(db, conversation, employee_id, content, msg_type, dify_conversation_id): return # 内容增强管线 enriched_content = await _step_enrich_image( db, content, msg_type, media_url, conversation_id, employee_id, ) enriched_content = await _enrich_with_last_ai_context( db, conversation_id, enriched_content, ) # 图谱短路 if await _step_graph_shortcut(db, conversation, employee_id, enriched_content, msg_type): return # 主推理(含 thinking 指示器 + 超时/无action 降级) result = await _step_call_dify( db, conversation, employee_id, conversation_id, content, enriched_content, dify_conversation_id, ) if result is None: return # 降级路径已全部处理(超时转人工 或 已推送降级卡片) # 后置处理管线 await _step_persist(db, conversation, employee_id, result) await _step_assets(db, employee_id, content, result) except Exception as e: import traceback logger.error(f"后台 AI 任务异常: {e}\n堆栈跟踪:\n{traceback.format_exc()}", exc_info=True) await _step_notify_failure(conversation_id, employee_id, "AI 服务异常,请转人工坐席或稍后重试。")