Fix: cannot parse images (#11044)

### What problem does this PR solve?

https://github.com/infiniflow/ragflow/issues/11043

### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)

rag/llm/cv_model.py

@@ -114,6 +114,28 @@ class Base(ABC):
 
         yield tk_count
 
+    @staticmethod
+    def image2base64_rawvalue(self, image):
+        # Return a base64 string without data URL header
+        if isinstance(image, bytes):
+            b64 = base64.b64encode(image).decode("utf-8")
+            return b64
+        if isinstance(image, BytesIO):
+            data = image.getvalue()
+            b64 = base64.b64encode(data).decode("utf-8")
+            return b64
+        with BytesIO() as buffered:
+            try:
+                image.save(buffered, format="JPEG")
+            except Exception:
+                 # reset buffer before saving PNG
+                buffered.seek(0)
+                buffered.truncate()
+                image.save(buffered, format="PNG")
+            data = buffered.getvalue()
+            b64 = base64.b64encode(data).decode("utf-8")
+        return b64
+
     @staticmethod
     def image2base64(image):
         # Return a data URL with the correct MIME to avoid provider mismatches
@@ -614,7 +636,15 @@ class GeminiCV(Base):
             if self.lang.lower() == "chinese"
             else "Please describe the content of this picture, like where, when, who, what happen. If it has number data, please extract them out."
         )
-        b64 = self.image2base64(image)
+
+        if image is bytes:
+            with BytesIO(image) as bio:
+                with open(bio) as img:
+                    input = [prompt, img]
+                    res = self.model.generate_content(input)
+                    return res.text, total_token_count_from_response(res)
+        else:
+            b64 = self.image2base64_rawvalue(image)
             with BytesIO(base64.b64decode(b64)) as bio:
                 with open(bio) as img:
                     input = [prompt, img]
@@ -623,9 +653,16 @@ class GeminiCV(Base):
 
     def describe_with_prompt(self, image, prompt=None):
         from PIL.Image import open
-
-        b64 = self.image2base64(image)
         vision_prompt = prompt if prompt else vision_llm_describe_prompt()
+
+        if image is bytes:
+            with BytesIO(image) as bio:
+                with open(bio) as img:
+                    input = [vision_prompt, img]
+                    res = self.model.generate_content(input)
+                    return res.text, total_token_count_from_response(res)
+        else:
+            b64 = self.image2base64_rawvalue(image)
             with BytesIO(base64.b64decode(b64)) as bio:
                 with open(bio) as img:
                     input = [vision_prompt, img]