Update app.py

app.py

@@ -106,20 +106,21 @@ def respond(message, api_key, max_tokens, top_p, temperature):
         if response.status_code == 200:
             assistant_reply = ""
             for line in response.iter_lines(decode_unicode=True):
-                print(f"Received line: {line}")  
+                if line.strip():
+                    if line.strip() == "[DONE]":
+                     break  
         
-                if line.strip(): 
-                    try:
-                        if line.startswith("data:"):
-                            line = line[5:].strip()
-                
-                        if line:  
-                            chunk = json.loads(line)  
-                            chunk_message = chunk.get("chunk", {}).get("content", "")
-                            assistant_reply += chunk_message
-                            yield assistant_reply
-                    except json.JSONDecodeError as e:
-                        print(f"Stream chunk error: {e} with line: {line}")
+              try:
+            if line.startswith("data:"):
+                line = line[5:].strip()
+
+            if line:
+                chunk = json.loads(line)
+                chunk_message = chunk.get("chunk", {}).get("content", "")
+                assistant_reply += chunk_message
+                yield assistant_reply
+        except json.JSONDecodeError as e:
+            print(f"Stream chunk error: {e} with line: {line}")
 
 
 
@@ -215,17 +216,14 @@ with gr.Blocks(css=css) as demo:
         last_message_state = gr.State("")  
 
         def user_interaction(message, history, api_key, max_tokens, top_p, temperature):
-            loading_message = history + [(message, "Loading...", "You", "P-ALPLE", sessions[api_key]["avatar"], ASSISTANT_PIC_PATH)]
-            yield render_message(loading_message), loading_message, ""
+             loading_message = history + [(message, "Loading...", "You", "P-ALPLE", sessions[api_key]["avatar"], ASSISTANT_PIC_PATH)]
+             yield render_message(loading_message), loading_message, ""
 
-            assistant_reply = ""
-            partial_history = history  
+            partial_history = history
             for partial_reply in respond(message, api_key, max_tokens, top_p, temperature):
-                partial_history = history + [(message, partial_reply, "You", "P-ALPLE", sessions[api_key]["avatar"], ASSISTANT_PIC_PATH)]
-                yield render_message(partial_history), partial_history, ""
+                 partial_history = history + [(message, partial_reply, "You", "P-ALPLE", sessions[api_key]["avatar"], ASSISTANT_PIC_PATH)]
+                 yield render_message(partial_history), partial_history, ""
 
-            history, assistant_reply = partial_history, partial_reply if 'partial_reply' in locals() else ("", "")
-            yield render_message(history), history, ""
 
         
         def regenerate_response(history, last_message, max_tokens, top_p, temperature):