fix:glm max tokens

app/api/v1/endpoints/image.py

@@ -62,20 +62,20 @@ async def process_image_ocr(
     try:
         log.info("Starting image OCR processing")
 
-        # Preprocess image
+        # Preprocess image (load only, no padding yet)
         preprocess_start = time.time()
         image = image_processor.preprocess(
             image_url=request.image_url,
             image_base64=request.image_base64,
         )
 
-        # Apply padding if enabled (before layout detection)
+        # Apply padding only for layout detection
         processed_image = image
         if image_processor and settings.is_padding:
             processed_image = image_processor.add_padding(image)
 
         preprocess_time = time.time() - preprocess_start
-        log.debug(f"Image preprocessing completed in {preprocess_time:.3f}s")
+        log.debug(f"Image loading completed in {preprocess_time:.3f}s")
 
         # Layout detection (using padded image if padding is enabled)
         layout_start = time.time()
@@ -83,14 +83,14 @@ async def process_image_ocr(
         layout_time = time.time() - layout_start
         log.info(f"Layout detection completed in {layout_time:.3f}s")
 
-        # OCR recognition
+        # OCR recognition (use original image without padding)
         ocr_start = time.time()
         if layout_info.MixedRecognition:
             recognition_method = "MixedRecognition (MinerU)"
             log.info(f"Using {recognition_method}")
 
-            # Convert numpy array to image bytes (image already padded)
-            success, encoded_image = cv2.imencode(".png", processed_image)
+            # Convert original image (without padding) to bytes
+            success, encoded_image = cv2.imencode(".png", image)
             if not success:
                 raise RuntimeError("Failed to encode image")
 
@@ -100,7 +100,29 @@ async def process_image_ocr(
         else:
             recognition_method = "FormulaOnly (GLMOCR)"
             log.info(f"Using {recognition_method}")
-            ocr_result = glmocr_service.recognize(processed_image)
+
+            # Try GLM-OCR first, fallback to MinerU if token limit exceeded
+            try:
+                ocr_result = glmocr_service.recognize(image)
+            except Exception as e:
+                error_msg = str(e)
+                # Check if error is due to token limit (max_model_len exceeded)
+                if "max_model_len" in error_msg or "decoder prompt" in error_msg or "BadRequestError" in error_msg:
+                    log.warning(f"GLM-OCR failed due to token limit: {error_msg}")
+                    log.info("Falling back to MinerU for recognition")
+                    recognition_method = "FormulaOnly (MinerU fallback)"
+
+                    # Convert original image to bytes for MinerU
+                    success, encoded_image = cv2.imencode(".png", image)
+                    if not success:
+                        raise RuntimeError("Failed to encode image")
+
+                    image_bytes = BytesIO(encoded_image.tobytes())
+                    image_bytes.seek(0)
+                    ocr_result = mineru_service.recognize(image_bytes)
+                else:
+                    # Re-raise other errors
+                    raise
         ocr_time = time.time() - ocr_start
 
         total_time = time.time() - preprocess_start

app/services/ocr_service.py

@@ -532,8 +532,10 @@ class GLMOCRService(OCRServiceBase):
 
         Returns:
             Dict with 'latex', 'markdown', 'mathml', 'mml' keys.
+        
+        Raises:
+            RuntimeError: If recognition fails (preserves original exception for fallback handling).
         """
-        try:
         # Add padding to image
         padded_image = self.image_processor.add_padding(image)
 
@@ -549,6 +551,7 @@ class GLMOCRService(OCRServiceBase):
         prompt = "Formula Recognition:"
         messages = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": image_url}}, {"type": "text", "text": prompt}]}]
 
+        # Don't catch exceptions here - let them propagate for fallback handling
         response = self.openai_client.chat.completions.create(
             model="glm-ocr",
             messages=messages,
@@ -574,8 +577,6 @@ class GLMOCRService(OCRServiceBase):
             "mml": convert_result.mml,
             "markdown": markdown_content,
         }
-        except Exception as e:
-            raise RuntimeError(f"GLM formula recognition failed: {e}") from e
 
     def recognize(self, image: np.ndarray) -> dict:
         """Recognize content using GLM-4V.