Fix response extraction - prevent truncation at beginning of JSON responses

gradio_app.py

@@ -261,12 +261,36 @@ def chat_with_model(message, history, temperature, json_mode=False, json_templat
         # Decode response
         generated_text = model_manager.tokenizer.decode(outputs[0], skip_special_tokens=True)
         
+        # Debug logging
+        logger.info(f"Full generated text length: {len(generated_text)} characters")
+        logger.info(f"Generated text preview: {generated_text[:300]}...")
+        logger.info(f"Generated text ending: ...{generated_text[-300:]}")
+        
         # Extract the response part (remove the prompt)
         if "<|start_header_id|>assistant<|end_header_id|>" in generated_text:
             response = generated_text.split("<|start_header_id|>assistant<|end_header_id|>")[-1].strip()
         else:
-            # Fallback: try to remove the prompt by length
-            response = generated_text[len(prompt):].strip()
+            # Improved fallback: look for common JSON starting patterns
+            response = generated_text
+            
+            # Try to find where the actual response starts
+            json_start_patterns = ['[', '{', '"']
+            for pattern in json_start_patterns:
+                if pattern in generated_text:
+                    # Find the first occurrence that looks like the start of JSON
+                    start_idx = generated_text.find(pattern)
+                    if start_idx > len(prompt) // 2:  # Make sure it's after the prompt
+                        response = generated_text[start_idx:].strip()
+                        break
+            
+            # Ultimate fallback: use the last portion of the text
+            if response == generated_text:
+                # Split by common delimiters and take the largest chunk
+                chunks = generated_text.split('\n\n')
+                if len(chunks) > 1:
+                    response = chunks[-1].strip()
+                else:
+                    response = generated_text[len(prompt)//2:].strip()
         
         # Log response length for debugging
         logger.info(f"Generated response length: {len(response)} characters")