Update app.py

app.py

@@ -9,9 +9,7 @@ client = OpenAI(
     base_url="https://api.runpod.ai/v2/vllm-k0g4c60zor9xuu/openai/v1",
 )
 
-def runpod_chat(question, history):
-    if not history:
-        history = []
+def runpod_chat(question, history=[]):
     history.append({"role": "user", "content": question})
 
     response_stream = client.chat.completions.create(
@@ -31,14 +29,13 @@ def runpod_chat(question, history):
         time.sleep(0.3)  # Simulate typing delay
         yield "RunPod: " + response
 
+    return history  # Return updated history to maintain state
+
 # Set up the Gradio interface
 iface = gr.Interface(
     fn=runpod_chat,
-    inputs=[
-        gr.Textbox(label="Enter your question:"),
-        gr.State()
-    ],
-    outputs="chat",
+    inputs=[gr.Textbox(label="Enter your question:"), gr.State()],
+    outputs=["chat", gr.State()],
     title="RunPod Chat",
     description="This app interfaces with RunPod's API to provide responses to your queries."
 )