app.py

import gradio as gr

def general_chat(text: str, gr_history: list):
    yield text

def online_chat(gr_webcam_image, gr_history: list):
    if gr_webcam_image:
        query = {"role": "user", "content": (gr_webcam_image, )} 
        response = {"role": "assistant", "content": 'this is a frame.'}
        gr_history.extend([query, response])
        yield gr_history

with gr.Blocks() as demo:
    gr.Markdown("## Simple Streaming Record and Response")
    with gr.Row():
        with gr.Column():
            gr_webcam_image = gr.Image(
                label='WebCam',
                sources="webcam",
                height=250,
                type='filepath')
        gr_chatinterface_ol = gr.ChatInterface(
            fn=general_chat,
            type="messages",
            multimodal=True,
        )
        gr_webcam_image.stream(
            fn=online_chat,
            inputs=[gr_webcam_image, gr_chatinterface_ol.chatbot],
            outputs=[gr_chatinterface_ol.chatbot],
            stream_every=1,
            concurrency_limit=30,
        )
demo.queue(default_concurrency_limit=100, max_size=100).launch(share=True,
                                                               max_threads=100,
                                                               ssr_mode=False)