feat: add image input normalization and enhance embedding logic for image URLs

apps/models_provider/base_model_provider.py

@@ -134,6 +134,23 @@ def embed_images(self, images: List[str]) -> List[List[float]]:
         """Embed data URLs or provider-supported image URLs."""
         raise AppApiException(500, _("The current embedding model does not support image embedding"))
 
+    @staticmethod
+    def normalize_image_input(url: str, keep_data_prefix: bool = False) -> str:
+        """归一化图片输入为上游接口可识别的字符串。
+
+        - http(s) URL：原样返回
+        - data:image/...;base64,xxx：keep_data_prefix=True 保留 data URI（阿里百炼等），
+          False 只返回 base64 内容（腾讯 TokenHub 等）
+        - 其它字符串：视为已是目标格式，原样返回
+        """
+        if url.startswith(("http://", "https://")):
+            return url
+        if url.startswith("data:") and ";base64," in url:
+            if keep_data_prefix:
+                return url
+            return url.split(";base64,", 1)[1]
+        return url
+
 
 class BaseModelCredential(ABC):
     @abstractmethod

apps/models_provider/impl/aliyun_bai_lian_model_provider/model/embedding.py

@@ -71,7 +71,9 @@ def embed_images(self, images: List[str]) -> List[List[float]]:
         """对图片 URL / data URL 做独立向量化（每张图生成一个向量）。"""
         if not self.supports_image_embedding():
             return []
-        return self._call_multimodal([{"image": image} for image in images])
+        return self._call_multimodal(
+            [{"image": self.normalize_image_input(image, keep_data_prefix=True)} for image in images]
+        )
 
     def _multimodal_base_url(self) -> str:
         """DashScope 原生多模态接口走 /api/v1，与 OpenAI 兼容地址区分开。"""

apps/models_provider/impl/tencent_model_provider/model/embedding.py

@@ -77,7 +77,19 @@ def embed_query(self, text: str) -> List[float]:
     def embed_images(self, images: List[str]) -> List[List[float]]:
         if not self.supports_image_embedding():
             return []
-        return [self._embed_multimodal([{"type": "image_url", "image_url": {"url": url}}]) for url in images]
+        return [
+            self._embed_multimodal([{"type": "image_url", "image_url": {"url": self._to_base64_content(url)}}])
+            for url in images
+        ]
+
+    @staticmethod
+    def _to_base64_content(url: str) -> str:
+        """TokenHub 的 image_url.url 接受 URL 或 base64 内容。
+
+        MaxKB 传入的图片是 data:image/...;base64,xxx 形式的 data URL，
+        这里剥掉 data: 前缀，转换为纯 base64 内容再交给接口。
+        """
+        return MaxKBBaseEmbeddingModel.normalize_image_input(url, keep_data_prefix=False)
 
     def _embed_multimodal(self, items: list) -> List[float]:
         payload = {"model": self.model_name, "input": items, "encoding_format": "float", **self.params}