fix: stabilize backend services and frontend pointer

fastapi_modules/fastapi_leaudit/services/impl/pageQualityServiceImpl.py

@@ -26,6 +26,18 @@ from fastapi_modules.fastapi_leaudit.services.impl.ossServiceImpl import OssServ
 
 logger = logging.getLogger(__name__)
 
+_PAGE_QUALITY_VLM_PROMPT = """
+你是文档扫描图片质量检测员。请判断这 1 页文档图片是否适合继续做 OCR 与合同/公文评查。
+
+判定标准：
+1. pass：文字主体清晰、方向正常、没有明显截断，能稳定阅读。
+2. review：存在轻微模糊、倾斜、阴影、低对比度、局部遮挡、轻微截断，建议人工确认但仍可能可读。
+3. reject：严重模糊、重影、过曝/过暗、页面大面积缺失、关键文字不可辨认、方向严重错误、空白页或非文档页，建议重拍。
+
+只输出 JSON，不要输出 Markdown，不要解释额外文本：
+{"status":"pass|review|reject","score":0.0到1.0,"reason":"20字以内中文原因"}
+""".strip()
+
 
 class PageQualityServiceImpl(IPageQualityService):
     """页级图片质量服务实现。"""
@@ -33,6 +45,7 @@ class PageQualityServiceImpl(IPageQualityService):
     def __init__(self) -> None:
         self.OssService = OssServiceImpl()
         self.DocumentService = None
+        self.VlmClient = None
 
     async def DispatchForDocument(
         self,
@@ -282,7 +295,7 @@ class PageQualityServiceImpl(IPageQualityService):
             reject_pages = 0
             async with GetAsyncSession() as session:
                 for page_num, page_image in page_images:
-                    status, score, reason = self._classify_page_image(page_image)
+                    status, score, reason = await self._classify_page_image_by_vlm(page_image)
                     if status == "review":
                         review_pages += 1
                     elif status == "reject":
@@ -466,13 +479,52 @@ class PageQualityServiceImpl(IPageQualityService):
         finally:
             doc.close()
 
-    def _classify_page_image(self, image_bytes: bytes) -> tuple[str, float, str | None]:
-        size = len(image_bytes)
-        if size < 25_000:
-            return "reject", 0.2, "页面图像内容过少或清晰度较低，建议重拍"
-        if size < 60_000:
-            return "review", 0.45, "页面疑似存在模糊，建议人工确认"
-        return "pass", 0.9, None
+    async def _classify_page_image_by_vlm(self, image_bytes: bytes) -> tuple[str, float, str | None]:
+        """使用 VLM 对单页图片做质量判定。VLM 异常不能默认为通过。"""
+        client = self._vlm_client()
+        if client is None:
+            return "review", 0.5, "VLM未配置，需人工确认图片质量"
+
+        try:
+            result = await client.extract_multifield(
+                prompt=_PAGE_QUALITY_VLM_PROMPT,
+                images_data_urls=[self._image_data_url(image_bytes)],
+                max_tokens=300,
+            )
+        except Exception as exc:
+            logger.warning("VLM page quality detection failed: %s", exc)
+            return "review", 0.5, "VLM图片质量检测失败，需人工确认"
+
+        status = str((result or {}).get("status") or "").strip().lower()
+        if status not in {"pass", "review", "reject"}:
+            return "review", 0.5, "VLM返回结果不可用，需人工确认"
+
+        score = self._normalize_quality_score((result or {}).get("score"), status)
+        reason = str((result or {}).get("reason") or "").strip() or None
+        if status != "pass" and not reason:
+            reason = "页面图片质量需人工确认"
+        return status, score, reason
+
+    def _vlm_client(self):
+        if self.VlmClient is None:
+            from fastapi_modules.fastapi_leaudit.leaudit_bridge.client_factory import create_vlm_client
+
+            self.VlmClient = create_vlm_client()
+        return self.VlmClient
+
+    def _image_data_url(self, image_bytes: bytes) -> str:
+        import base64
+
+        encoded = base64.b64encode(image_bytes).decode()
+        return f"data:image/png;base64,{encoded}"
+
+    def _normalize_quality_score(self, raw_score: Any, status: str) -> float:
+        defaults = {"pass": 0.9, "review": 0.5, "reject": 0.2}
+        try:
+            score = float(raw_score)
+        except (TypeError, ValueError):
+            return defaults[status]
+        return max(0.0, min(1.0, score))
 
     def _document_service(self):
         if self.DocumentService is None: