From ac005c585a1d9cb981952d244ae812e1e330ff80 Mon Sep 17 00:00:00 2001
From: wxg0103 <727495428@qq.com>
Date: Thu, 3 Sep 2026 15:34:29 +0800
Subject: [PATCH] feat: enhance multimodal embedding support and refactor
 embedding logic

---
 .../model/embedding.py                        | 55 +++++++++++++------
 1 file changed, 37 insertions(+), 18 deletions(-)

diff --git a/apps/models_provider/impl/aliyun_bai_lian_model_provider/model/embedding.py b/apps/models_provider/impl/aliyun_bai_lian_model_provider/model/embedding.py
index 3f335bb466a..c38bb790756 100644
--- a/apps/models_provider/impl/aliyun_bai_lian_model_provider/model/embedding.py
+++ b/apps/models_provider/impl/aliyun_bai_lian_model_provider/model/embedding.py
@@ -17,9 +17,6 @@
 
 
 class AliyunBaiLianEmbedding(MaxKBBaseEmbeddingModel):
-    def supports_image_embedding(self) -> bool:
-        return False
-
     model_name: str
     optional_params: dict
     api_base: str
@@ -35,6 +32,14 @@ def __init__(self, api_key, model_name: str, api_base: str, optional_params: dic
     def is_cache_model(self):
         return False
 
+    @staticmethod
+    def _is_multimodal(model_name: str) -> bool:
+        """判断模型是否为多模态向量模型（支持图片/视频独立向量）。"""
+        return any(k in model_name for k in ("vl-embedding", "embedding-vision", "multimodal"))
+
+    def supports_image_embedding(self) -> bool:
+        return self._is_multimodal(self.model_name)
+
     @staticmethod
     def new_instance(model_type, model_name, model_credential: Dict[str, object], **model_kwargs):
         optional_params = MaxKBBaseEmbeddingModel.filter_optional_params(model_kwargs)
@@ -51,21 +56,8 @@ def embed_query(self, text: str):
 
     def embed_documents(self, texts: List[str], chunk_size: int | None = None) -> List[List[float]]:
         # 处理多模态的向量化
-        if any(k in self.model_name for k in ("vl-embedding", "embedding-vision", "multimodal")):
-            dashscope.api_key = self.api_key
-            dashscope.base_http_api_url = self.api_base
-            multimodal_input = [{"text": text} for text in texts]
-            resp = dashscope.MultiModalEmbedding.call(
-                model=self.model_name,
-                input=multimodal_input,  # type: ignore
-                **self.optional_params,
-            )
-
-            if resp.status_code == HTTPStatus.OK:
-                embeddings_data = resp.output.get("embeddings", [])
-                return [item.get("embedding", []) for item in embeddings_data]
-            else:
-                raise Exception(f"MultiModalEmbedding call failed: status={resp.status_code}, message={resp.message}")
+        if self._is_multimodal(self.model_name):
+            return self._call_multimodal([{"text": text} for text in texts])
 
         if len(self.optional_params) > 0:
             res = self.client.create(
@@ -74,3 +66,30 @@ def embed_documents(self, texts: List[str], chunk_size: int | None = None) -> Li
         else:
             res = self.client.create(input=texts, model=self.model_name, encoding_format="float")
         return [e.embedding for e in res.data]
+
+    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])
+
+    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。"""
+        base = self.api_base or "https://dashscope.aliyuncs.com/api/v1"
+        if "/compatible-mode/" in base:
+            return base.split("/compatible-mode/")[0] + "/api/v1"
+        return base
+
+    def _call_multimodal(self, items: List[dict]) -> List[List[float]]:
+        dashscope.api_key = self.api_key
+        dashscope.base_http_api_url = self._multimodal_base_url()
+        resp = dashscope.MultiModalEmbedding.call(
+            model=self.model_name,
+            input=items,  # type: ignore
+            **self.optional_params,
+        )
+
+        if resp.status_code == HTTPStatus.OK:
+            embeddings_data = resp.output.get("embeddings", [])
+            return [item.get("embedding", []) for item in embeddings_data]
+        raise Exception(f"MultiModalEmbedding call failed: status={resp.status_code}, message={resp.message}")