Update app.py

app.py

@@ -135,35 +135,43 @@ model = ModelWrapper(MODEL_NAME)
 
 # — Gradio inference function —
 def compliance_check(rules_text, transcript_text, thinking):
+    # This is the general byte limit for any response leaving this function
+    MAX_RESPONSE_BYTES = 4096 
+
     try:
         rules = [r for r in rules_text.split("\n") if r.strip()]
         inp = format_rules(rules) + format_transcript(transcript_text)
         
-        # Limit max tokens to prevent oversized responses
         out = model.get_response(inp, enable_thinking=thinking, max_new_tokens=256)
         
-        # Clean up any malformed output and ensure it's a string
+        # Clean up the successful output
         out = str(out).strip()
         if not out:
             out = "No response generated. Please try with different input."
-
-        # Ensure the response isn't too long for an HTTP response by checking byte length
-        max_bytes = 2500  # A more generous limit, in bytes
-        out_bytes = out.encode('utf-8')
-
-        if len(out_bytes) > max_bytes:
-            # Truncate the byte string, then decode back to a string, ignoring errors
-            # This prevents cutting a multi-byte character in half
-            truncated_bytes = out_bytes[:max_bytes]
-            out = truncated_bytes.decode('utf-8', errors='ignore')
-            out += "\n\n[Response truncated to prevent server errors]"
             
-        return out
-        
     except Exception as e:
-        error_msg = f"Error: {str(e)[:200]}"  # Limit error message length
-        print(f"Full error: {e}")
-        return error_msg
+        # If any error happens, create an error message instead
+        full_error = str(e)
+        print(f"Full error: {full_error}") # Log the full error for debugging
+        # The output 'out' will now be this error message
+        out = f"An error occurred: {full_error}"
+
+    # --- UNIVERSAL BYTE CHECK FOR ALL OUTPUTS ---
+    # This block now runs for both successful results and error messages.
+    
+    # Encode the final output (whether result or error) to bytes
+    out_bytes = out.encode('utf-8')
+
+    # Check if the byte length exceeds our limit
+    if len(out_bytes) > MAX_RESPONSE_BYTES:
+        # Truncate the byte string safely
+        truncated_bytes = out_bytes[:MAX_RESPONSE_BYTES]
+        out = truncated_bytes.decode('utf-8', errors='ignore')
+        # Add a clear message indicating it was truncated
+        out += "\n\n[Response truncated to prevent server errors]"
+            
+    return out
+
 
 
 # — build Gradio interface —