app.py

import os
import re
import gradio as gr
import edge_tts
import asyncio
import time
import tempfile
from huggingface_hub import InferenceClient

DESCRIPTION = """ # <center><b>ZARVIS⚡</b></center>
        ### <center>A personal voice assistant for YOU
        ### <center>I'm your ZEN Voice Assistant.</center>
        """

MORE = """ ## TRY Other Models
        ### https://zenai.biz
        """

Fast = """## Fastest Model"""
Complex = """## Best in Complex Question"""
Detail = """## Best for Detailed Generation or Long Answers"""

# ----------------- MODEL 1 ----------------- #
# Mistral-based model
client1 = InferenceClient("mistralai/Mixtral-8x7B-Instruct-v0.1")

system_instructions1 = (
    "[SYSTEM] Answer as Real ZARVIS, made by 'ZEN'. "
    "Keep conversation very short, clear, friendly, and concise. "
    "The text provided is a request for a specific type of response from you, the virtual assistant. "
    "The request asks you to provide friendly responses as if you are the character ZARVIS, made by Tony Stark. "
    "Avoid introductions and start answering the query directly, elaborating on all aspects. "
    "As an AI-powered assistant, your task is to generate responses that appear as if they are created by the user, "
    "not an AI-powered assistant. [USER]"
)

async def generate1(prompt):
    generate_kwargs = dict(
        temperature=0.6,
        max_new_tokens=756,
        top_p=0.95,
        repetition_penalty=1,
        do_sample=True,
        seed=42,
    )
    formatted_prompt = system_instructions1 + prompt + "[ZARVIS]"
    stream = client1.text_generation(
        formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True
    )
    output = ""
    for response in stream:
        output += response.token.text

    communicate = edge_tts.Communicate(output)
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
        tmp_path = tmp_file.name
        await communicate.save(tmp_path)
    yield tmp_path

# ----------------- MODEL 2 ----------------- #
# Llama-based model for more complex tasks
client2 = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")

system_instructions2 = (
    "[SYSTEM] Answer as Real ZARVIS, made by 'ZEN'. "
    "You must answer in a friendly style and easy manner. "
    "You can answer complex questions. "
    "Do not say who you are or greet; simply start answering. "
    "Stop as soon as you have given the complete answer. [USER]"
)

async def generate2(prompt):
    generate_kwargs = dict(
        temperature=0.6,
        max_new_tokens=512,
        top_p=0.95,
        repetition_penalty=1,
        do_sample=True,
    )    
    formatted_prompt = system_instructions2 + prompt + "[ASSISTANT]"
    stream = client2.text_generation(
        formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True
    )
    output = ""
    for response in stream:
        output += response.token.text

    communicate = edge_tts.Communicate(output)
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
        tmp_path = tmp_file.name
        await communicate.save(tmp_path)
    yield tmp_path

# ----------------- MODEL 3 ----------------- #
# Another Llama-based model for longer, more detailed answers
client3 = InferenceClient("meta-llama/Meta-Llama-3-70B-Instruct")

system_instructions3 = (
    "[SYSTEM] The text provided is a request for a specific type of response from me, the virtual assistant. "
    "I should provide detailed and friendly responses as if I am the character ZARVIS, inspired by Tony Stark. "
    "Avoid introductions and start answering the query directly, elaborating on all aspects of the request. "
    "As an AI-powered assistant, my task is to generate responses that appear as if they are created by the user, "
    "not an AI-powered assistant. [USER]"
)

async def generate3(prompt):
    generate_kwargs = dict(
        temperature=0.6,
        max_new_tokens=2048,
        top_p=0.95,
        repetition_penalty=1,
        do_sample=True,
    )    
    formatted_prompt = system_instructions3 + prompt + "[ASSISTANT]"
    stream = client3.text_generation(
        formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=True
    )
    output = ""
    for response in stream:
        output += response.token.text

    communicate = edge_tts.Communicate(output)
    with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as tmp_file:
        tmp_path = tmp_file.name
        await communicate.save(tmp_path)
    yield tmp_path

# ----------------- Gradio Interface ----------------- #
with gr.Blocks(css="style.css") as demo:
    gr.Markdown(DESCRIPTION)
    with gr.Row():
        user_input = gr.Textbox(label="Prompt", value="What is Wikipedia")
        input_text = gr.Textbox(label="(Optional) Additional Input", elem_id="important")
        output_audio = gr.Audio(
            label="ZARVIS",
            type="filepath",
            interactive=False,
            autoplay=True,
            elem_classes="audio"
        )
    with gr.Row():
        translate_btn = gr.Button("Response")
        translate_btn.click(
            fn=generate1, 
            inputs=user_input, 
            outputs=output_audio, 
            api_name="translate"
        )

    gr.Markdown(MORE)

if __name__ == "__main__":
    demo.queue(max_size=200).launch()