debase 0.4.1__py3-none-any.whl → 0.4.2__py3-none-any.whl

debase/reaction_info_extractor.py

@@ -1332,12 +1332,28 @@ class ReactionExtractor:
             y_offset += pix.height * scale
             
         # Convert the page to a pixmap
-        mat = fitz.Matrix(5.0, 5.0)  # 5x zoom for quality
+        # Limit zoom factor to avoid creating excessively large images
+        # Gemini has limits on image size (approx 20MB or 20 megapixels)
+        zoom = 5.0
+        estimated_pixels = (max_width * zoom) * (total_height * zoom)
+        max_pixels = 20_000_000  # 20 megapixels
+        
+        if estimated_pixels > max_pixels:
+            # Calculate appropriate zoom to stay under limit
+            zoom = min(5.0, (max_pixels / (max_width * total_height)) ** 0.5)
+            LOGGER.warning(f"Reducing zoom from 5.0 to {zoom:.2f} to stay under {max_pixels/1e6:.1f} megapixel limit")
+        
+        mat = fitz.Matrix(zoom, zoom)
         combined_pix = page.get_pixmap(matrix=mat)
         combined_pix = self._ensure_rgb_pixmap(combined_pix)
         
         # Convert to PNG and return
         img_bytes = combined_pix.tobytes("png")
+        
+        # Check final size
+        final_size_mb = len(img_bytes) / (1024 * 1024)
+        if final_size_mb > 20:
+            LOGGER.warning(f"Combined image is {final_size_mb:.1f}MB, may be too large for vision API")
         output_doc.close()
         
         # Save debug file if available
@@ -2317,39 +2333,51 @@ Different campaigns may use different model reactions and substrates.
                         }
                     )
                     
-                    response = model.generate_content(content_parts)
-                    
-                    # Track token usage if available
                     try:
-                        if hasattr(response, 'usage_metadata'):
-                            input_tokens = getattr(response.usage_metadata, 'prompt_token_count', 0)
-                            output_tokens = getattr(response.usage_metadata, 'candidates_token_count', 0)
-                            if input_tokens or output_tokens:
-                                try:
-                                    from .wrapper import add_token_usage
-                                    add_token_usage('reaction_info_extractor', input_tokens, output_tokens)
-                                except ImportError:
-                                    pass  # wrapper not available
-                    except Exception:
-                        pass  # token tracking is best-effort
-                    
-                    # Parse JSON from response
-                    if response and response.text:
-                        # Save debug output
-                        if self.debug_dir:
-                            timestamp = int(time.time())
-                            _dump(prompt, self.debug_dir / f"model_reaction_multimodal_prompt_{timestamp}.txt")
-                            _dump(response.text, self.debug_dir / f"model_reaction_multimodal_response_{timestamp}.txt")
+                        response = model.generate_content(content_parts)
                         
-                        # Extract JSON from response
-                        text = response.text.strip()
-                        if text.startswith("```json"):
-                            text = text[7:]
-                        if text.endswith("```"):
-                            text = text[:-3]
-                        data = json.loads(text.strip())
-                    else:
-                        raise ValueError("Empty response from multimodal model")
+                        # Track token usage if available
+                        try:
+                            if hasattr(response, 'usage_metadata'):
+                                input_tokens = getattr(response.usage_metadata, 'prompt_token_count', 0)
+                                output_tokens = getattr(response.usage_metadata, 'candidates_token_count', 0)
+                                if input_tokens or output_tokens:
+                                    try:
+                                        from .wrapper import add_token_usage
+                                        add_token_usage('reaction_info_extractor', input_tokens, output_tokens)
+                                    except ImportError:
+                                        pass  # wrapper not available
+                        except Exception:
+                            pass  # token tracking is best-effort
+                        
+                        # Parse JSON from response
+                        if response and response.text:
+                            # Save debug output
+                            if self.debug_dir:
+                                timestamp = int(time.time())
+                                _dump(prompt, self.debug_dir / f"model_reaction_multimodal_prompt_{timestamp}.txt")
+                                _dump(response.text, self.debug_dir / f"model_reaction_multimodal_response_{timestamp}.txt")
+                            
+                            # Extract JSON from response
+                            text = response.text.strip()
+                            if text.startswith("```json"):
+                                text = text[7:]
+                            if text.endswith("```"):
+                                text = text[:-3]
+                            data = json.loads(text.strip())
+                        else:
+                            raise ValueError("Empty response from multimodal model")
+                    except Exception as vision_error:
+                        LOGGER.error("Vision API call failed: %s", vision_error)
+                        LOGGER.info("Falling back to text-only extraction")
+                        # Fall back to text-only extraction
+                        data = generate_json_with_retry(
+                            self.model,
+                            prompt,
+                            temperature=self.cfg.model_reaction_temperature,
+                            debug_dir=self.debug_dir,
+                            tag="model_reaction_fallback"
+                        )
                 else:
                     # Fall back to text-only extraction
                     data = generate_json_with_retry(