feat(backend): knowledge iteration + vision + neo4j + response contract source

dependencies.py 拆分为 dependencies/ 包; 新增 vision/ragflow_ingestion/neo4j 客户端与 h5_ai_task; alembic 045 图置信度迁移; 响应契约统一收尾。

backend/app/services/ai_service.py

@@ -199,36 +199,96 @@ class AIService:
         conversation_id: Optional[str] = None,
         user_id: Optional[str] = None,
     ) -> AsyncGenerator[Dict[str, Any], None]:
-        """调用 Dify API 获取流式 AI 回复（SSE）。
-
-        Args:
-            message: 员工发送的消息内容
-            conversation_id: Dify 会话ID
-            user_id: 员工企微 UserID
+        """调用 Dify API 获取流式 AI 回复（SSE），逐块 yield 给调用方。
 
         Yields:
-            Dict: {
-                "delta": str,        # 增量内容
-                "finished": bool,     # 是否结束
-                "conversation_id": str,
-                "hit": bool,         # 最终判断是否命中
-            }
+            Dict: {"delta": str, "finished": bool, "conversation_id": str, "hit": bool|None}
+            - 流式中间块：{"delta": 增量, "finished": False, "hit": None}
+            - 终态块：{"delta": "", "finished": True, "hit": 命中判断}
 
-        做什么：SSE 流式读取 Dify 返回，逐块 yield 给调用方
-        为什么：
-            - 流式返回能提升用户体验（不用等 AI 全部生成完才显示）
-            - 通过 WebSocket 推送增量内容到 H5 前端
-            - 目前第一步先实现非流式，流式作为后续优化
+        实现：
+            - stream=True 走 SSE，解析 data: {...} 行，逐块 yield delta
+            - 流结束后用完整内容整体判断 hit（_check_knowledge_hit）
+        容错：若 Dify 不支持流式 / 超时 / 非 SSE 格式，catch 后 fallback 到
+              get_reply 非流式，yield 一次完整内容（前端退化为"整段到达"，
+              功能不破，仅无逐字动画）。
         """
-        # TODO: 第一步简化，先 yield 完整内容（非真正流式）
-        # 后续优化：解析 SSE 事件流，逐块 yield
-        result = await self.get_reply(message, conversation_id, user_id)
-        yield {
-            "delta": result["content"],
-            "finished": True,
-            "conversation_id": result["conversation_id"],
-            "hit": result["hit"],
+        payload = {
+            "model": "Chat",
+            "messages": [{"role": "user", "content": message}],
+            "stream": True,
+            "temperature": 0.1,
         }
+        if conversation_id:
+            payload["conversation_id"] = conversation_id
+        if user_id:
+            payload["user"] = user_id
+
+        try:
+            client = await self._get_client()
+            full_parts: list = []
+            dify_conv_id = conversation_id or ""
+            async with client.stream("POST", self.api_url, json=payload) as response:
+                response.raise_for_status()
+                async for line in response.aiter_lines():
+                    if not line:
+                        continue
+                    line = line.strip()
+                    if not line.startswith("data:"):
+                        continue
+                    data = line[5:].strip()
+                    if data == "[DONE]":
+                        break
+                    try:
+                        chunk = json.loads(data)
+                    except json.JSONDecodeError:
+                        continue
+                    # OpenAI / Dify SSE 格式：choices[0].delta.content
+                    try:
+                        delta = chunk["choices"][0]["delta"].get("content", "")
+                    except (KeyError, IndexError, TypeError):
+                        delta = ""
+                    if delta:
+                        full_parts.append(delta)
+                        yield {
+                            "delta": delta,
+                            "finished": False,
+                            "conversation_id": dify_conv_id,
+                            "hit": None,
+                        }
+                    # Dify 可能在流式块里给出 conversation_id
+                    cid = chunk.get("conversation_id")
+                    if cid:
+                        dify_conv_id = cid
+
+            # 流结束：用完整内容判断命中
+            full_content = "".join(full_parts)
+            hit = self._check_knowledge_hit(full_content) if full_content else False
+            yield {
+                "delta": "",
+                "finished": True,
+                "conversation_id": dify_conv_id,
+                "hit": hit,
+            }
+        except Exception as e:
+            # 流式不可用（dify2openai 不支持 / 超时 / 非 SSE），回退非流式
+            logger.warning(f"Dify 流式失败，回退非流式: {e}")
+            try:
+                result = await self.get_reply(message, conversation_id, user_id)
+                yield {
+                    "delta": result["content"],
+                    "finished": True,
+                    "conversation_id": result["conversation_id"],
+                    "hit": result["hit"],
+                }
+            except Exception as e2:
+                logger.error(f"Dify 流式与非流式均失败: {e2}")
+                yield {
+                    "delta": "⚠️ AI 服务异常，请输入「IT」转人工或稍后重试。",
+                    "finished": True,
+                    "conversation_id": conversation_id or "",
+                    "hit": False,
+                }
 
     # --------------------------------------------------------------------------
     # 判断是否命中知识库