app.py

import gradio as gr
from huggingface_hub import InferenceClient
import random
import textwrap

# Define the model to be used
model = "mistralai/Mixtral-8x7B-Instruct-v0.1"
client = InferenceClient(model)

# Embedded system prompt
system_prompt_text = "You are a smart and helpful co-worker of Thailand based multi-national company PTT, and PTTEP. You help with any kind of request and provide a detailed answer to the question. But if you are asked about something unethical or dangerous, you must refuse and provide a safe and respectful way to handle that."

# Read the content of the info.md and info2.md files
with open("info.md", "r") as file:
    info_md_content = file.read()

with open("info2.md", "r") as file:
    info2_md_content = file.read()

# Chunk the info.md and info2.md content into smaller sections
chunk_size = 1500  # Adjust this size as needed to fit the context window
info_md_chunks = textwrap.wrap(info_md_content, chunk_size)
info2_md_chunks = textwrap.wrap(info2_md_content, chunk_size)

# Combine both sets of chunks
all_chunks = info_md_chunks + info2_md_chunks

# Function to initialize the conversation history with chunks
def initialize_history(chunks):
    history = []
    for chunk in chunks:
        history.append(("System Information", chunk))
    return history

# Initialize history with initial chunks
history = initialize_history(all_chunks[:2])  # Starting with the first two chunks for example

def format_prompt_mixtral(message, history):
    prompt = "<s>"
    prompt += f"{system_prompt_text}\n\n"  # Add the system prompt

    # Include the initial context from the chunks
    for user_prompt, bot_response in history:
        prompt += f"[INST] {user_prompt} [/INST] {bot_response}</s> "

    # Add the current user message
    prompt += f"[INST] {message} [/INST]"
    return prompt

def chat_inf(message, history, seed, temp, tokens, top_p, rep_p):
    generate_kwargs = dict(
        temperature=temp,
        max_new_tokens=tokens,
        top_p=top_p,
        repetition_penalty=rep_p,
        do_sample=True,
        seed=seed,
    )

    formatted_prompt = format_prompt_mixtral(message, history)
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
    output = ""
    for response in stream:
        output += response.token.text
        yield [(message, output)]
    history.append((message, output))
    yield history

def clear_fn():
    global history
    history = initialize_history(all_chunks[:2])  # Reset to initial chunks
    return None, None

rand_val = random.randint(1, 1111111111111111)

def check_rand(rand, val):
    if rand:
        return random.randint(1, 1111111111111111)
    else:
        return int(val)

with gr.Blocks() as app:
    gr.HTML("""<center><h1 style='font-size:xx-large;'>PTT Chatbot</h1><br><h3>running on Huggingface Inference </h3><br><h7>EXPERIMENTAL</center>""")
    with gr.Row():
        chat = gr.Chatbot(height=500)
    with gr.Group():
        with gr.Row():
            with gr.Column(scale=3):
                inp = gr.Textbox(label="Prompt", lines=5, interactive=True)  # Increased lines and interactive
                with gr.Row():
                    with gr.Column(scale=2):
                        btn = gr.Button("Chat")
                    with gr.Column(scale=1):
                        with gr.Group():
                            stop_btn = gr.Button("Stop")
                            clear_btn = gr.Button("Clear")
            with gr.Column(scale=1):
                with gr.Group():
                    rand = gr.Checkbox(label="Random Seed", value=True)
                    seed = gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, step=1, value=rand_val)
                    tokens = gr.Slider(label="Max new tokens", value=3840, minimum=0, maximum=8000, step=64, interactive=True, visible=True, info="The maximum number of tokens")
                    temp = gr.Slider(label="Temperature", step=0.01, minimum=0.01, maximum=1.0, value=0.9)
                    top_p = gr.Slider(label="Top-P", step=0.01, minimum=0.01, maximum=1.0, value=0.9)
                    rep_p = gr.Slider(label="Repetition Penalty", step=0.1, minimum=0.1, maximum=2.0, value=1.0)

    hid1 = gr.Number(value=1, visible=False)

    def on_chat(message, chat, seed, temp, tokens, top_p, rep_p):
        chat.clear()
        history = initialize_history(all_chunks[:2])
        for response in chat_inf(message, history, seed, temp, tokens, top_p, rep_p):
            chat.append(*response)
    
    go = btn.click(on_chat, [inp, chat, seed, temp, tokens, top_p, rep_p], chat)

    stop_btn.click(None, None, None, cancels=[go])
    clear_btn.click(clear_fn, None, [inp, chat])

app.queue(default_concurrency_limit=10).launch(share=True, auth=("admin", "0112358"))