Update app.py

app.py

@@ -89,7 +89,7 @@ class ModelWrapper:
               prompt = self.tokenizer.apply_chat_template(message, tokenize=False, continue_final_message=True, enable_thinking=False)
         return prompt
 
-    def get_response(self, input, temperature=0.7, top_k=20, top_p=0.8, max_new_tokens=128, enable_thinking=True, system_prompt=SYSTEM_PROMPT):
+    def get_response(self, input, temperature=0.7, top_k=20, top_p=0.8, max_new_tokens=256, enable_thinking=True, system_prompt=SYSTEM_PROMPT):
         """Generate and decode the response with the recommended temperature settings for thinking and non-thinking."""
         print("Generating response...")
         
@@ -136,55 +136,50 @@ model = ModelWrapper(MODEL_NAME)
 # — Gradio inference function —
 def compliance_check(rules_text, transcript_text, thinking):
     try:
-        # Extremely conservative limits to prevent HTTP errors
-        if len(rules_text) > 500:
-            return "Error: Rules text too long. Please limit to 500 characters."
-        
-        if len(transcript_text) > 1000:
-            return "Error: Transcript too long. Please limit to 1000 characters."
-        
         rules = [r for r in rules_text.split("\n") if r.strip()]
         inp = format_rules(rules) + format_transcript(transcript_text)
         
-        # Very small token limit
-        out = model.get_response(inp, enable_thinking=thinking, max_new_tokens=128)
-        
-        # Aggressive truncation - keep response very short
-        max_chars = 800  # Very conservative
-        if len(out) > max_chars:
-            out = out[:max_chars] + "..."
-        
-        # Ensure it's valid UTF-8 and clean
-        out = out.encode('utf-8', errors='ignore').decode('utf-8').strip()
+        # 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
+        out = str(out).strip()
         if not out:
-            out = "No response generated."
+            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:
-        # Very short error message
-        return f"Error: Processing failed"
+        error_msg = f"Error: {str(e)[:200]}"  # Limit error message length
+        print(f"Full error: {e}")
+        return error_msg
+
 
 # — build Gradio interface —
 demo = gr.Interface(
     fn=compliance_check,
     inputs=[
-        gr.Textbox(lines=3, label="Rules (one per line)", max_lines=5, placeholder="Enter 1-3 rules, keep short"),
-        gr.Textbox(lines=5, label="Transcript", max_lines=8, placeholder="Enter transcript, max 1000 chars"),
-        gr.Checkbox(label="Enable ⟨think⟩ mode", value=False)  # Default to False for faster processing
+        gr.Textbox(lines=5, label="Rules (one per line)", max_lines=10),
+        gr.Textbox(lines=10, label="Transcript", max_lines=15),
+        gr.Checkbox(label="Enable ⟨think⟩ mode", value=True)
     ],
-    outputs=gr.Textbox(label="Compliance Output", lines=8, max_lines=10),
+    outputs=gr.Textbox(label="Compliance Output", lines=10, max_lines=15),
     title="DynaGuard Compliance Checker",
-    description="Short inputs only. Rules: max 500 chars. Transcript: max 1000 chars.",
-    flagging_mode="never",
+    description="Paste your rules & transcript, then hit Submit.",
+    allow_flagging="never",
     show_progress=True
 )
 
 if __name__ == "__main__":
-    # Minimal launch configuration
-    demo.launch(
-        show_error=False,  # Hide detailed errors
-        quiet=True,        # Reduce logging
-        share=False
-    )
+    demo.launch()