diff --git a/.gitignore b/.gitignore
index ea77a85..fbff7d3 100644
--- a/.gitignore
+++ b/.gitignore
@@ -101,4 +101,5 @@ Thumbs.db
 # 本地导出的 IDE / 对话历史等（按需）
 history.txt
 .cursor/
-backend/.env.docker
\ No newline at end of file
+backend/.env.docker
+backend/uploads/
\ No newline at end of file
diff --git a/backend/services/vector_service.py b/backend/services/vector_service.py
index 06604e2..f780993 100644
--- a/backend/services/vector_service.py
+++ b/backend/services/vector_service.py
@@ -370,6 +370,68 @@ class VectorService:
         except Exception as e:
             logger.error(f"图片 OCR 处理失败: {e}")
             raise
+
+    async def _process_image_with_vision_fallback(self, file_path: str) -> List:
+        """
+        处理知识库图片：优先 OCR；OCR 无文字或失败时，必须使用视觉模型提取内容。
+        """
+        from langchain_core.documents import Document
+        from services.vision_service import VisionService
+
+        ocr_docs: List = []
+        try:
+            ocr_docs = await asyncio.to_thread(self._process_image_ocr, file_path)
+        except Exception as e:
+            logger.warning(f"图片 OCR 异常，将尝试视觉模型: {e}")
+
+        if ocr_docs:
+            return ocr_docs
+
+        logger.info("OCR 未识别到文字，使用视觉模型处理图片...")
+        vision_prompt = (
+            "详细描述图片中的内容：场景、人物、物体、图表及所有可见文字（逐字提取）。"
+            "用通顺中文输出，便于后续检索与问答。"
+        )
+
+        try:
+            with open(file_path, "rb") as f:
+                image_bytes = f.read()
+        except OSError as e:
+            logger.error(f"读取图片文件失败，无法进行视觉理解: {e}")
+            return []
+
+        ext = os.path.splitext(file_path)[1].lower()
+        mime_map = {
+            ".png": "image/png",
+            ".jpg": "image/jpeg",
+            ".jpeg": "image/jpeg",
+            ".bmp": "image/bmp",
+            ".webp": "image/webp",
+            ".gif": "image/gif",
+        }
+        vision_text = await VisionService.get_image_description_from_bytes(
+            image_bytes,
+            prompt=vision_prompt,
+            mime_hint=mime_map.get(ext, "image/jpeg"),
+        )
+
+        if not vision_text or not vision_text.strip():
+            logger.warning(f"视觉模型也未提取到内容: {file_path}")
+            return []
+
+        logger.info(f"视觉模型成功提取图片内容，共 {len(vision_text)} 字符")
+        return [
+            Document(
+                page_content=f"【图片内容描述】\n{vision_text.strip()}",
+                metadata={
+                    "source": file_path,
+                    "file_type": "image",
+                    "has_ocr": False,
+                    "has_vision": True,
+                    "vision_provider": "qwen-vl",
+                },
+            )
+        ]
     
     def _extract_images_from_docx(self, docx_path: str) -> List[str]:
         """
@@ -885,8 +947,8 @@ class VectorService:
             
             # 2. 加载文档（特殊处理图片 OCR 和 DOCX，放到线程池执行）
             if loader == "image_ocr":
-                logger.info("🔄 在线程池中执行图片 OCR...")
-                docs = await asyncio.to_thread(self._process_image_ocr, file_path)
+                logger.info("🔄 处理图片：OCR → 视觉模型回退...")
+                docs = await self._process_image_with_vision_fallback(file_path)
             elif loader == "docx_with_images":
                 logger.info("🔄 在线程池中处理 DOCX 文件（提取图片并 OCR）...")
                 docs, _ = await asyncio.to_thread(self._process_docx_with_images, file_path)  # 忽略图片路径（知识库暂不使用视觉模型）
@@ -913,7 +975,10 @@ class VectorService:
             logger.info(f"文档加载完成，共 {len(docs)} 个文档片段")
             
             if not docs:
-                error_msg = "未能从文件加载到任何内容"
+                error_msg = (
+                    "未能从文件加载到任何内容。"
+                    "图片请确认已配置 DASHSCOPE_API_KEY（视觉模型）和/或阿里云 OCR。"
+                )
                 logger.warning(error_msg)
                 return ProcessResult(success=False, chunks=[], chunk_count=0, error_message=error_msg)