app.py

import gradio as gr
from huggingface_hub import InferenceClient
from transformers import pipeline

# ---------------- CONFIG ----------------
MODEL_REPO = "HuggingFaceH4/zephyr-7b-beta"
TRANSLATOR_MODEL = "facebook/m2m100_418M"  # multilingual translator

SYSTEM_PROMPT_DEFAULT = (
    "You are Zephyr, a concise and polite AI assistant. "
    "Always respond in a formal tone and provide only the direct answer unless the user requests more detail."
)

MAX_NEW_TOKENS_DEFAULT = 512  # increased to handle long answers
TEMP_DEFAULT = 0.7
TOP_P_DEFAULT = 0.95
MAX_HISTORY_MESSAGES = 10  # limit chat history to prevent repetition

# Clients
client = InferenceClient(MODEL_REPO)
translator = pipeline("translation", model=TRANSLATOR_MODEL)


# ---------------- HELPERS ----------------
def is_translation_request(message: str) -> bool:
    triggers = ["translate", "traduce", "ترجم", "traduire", "übersetze"]
    if any(t in message.lower() for t in triggers):
        return True
    non_ascii_ratio = sum(1 for c in message if ord(c) > 127) / max(len(message), 1)
    return non_ascii_ratio > 0.4


# ---------------- CHAT FUNCTION ----------------
def stream_response(message, chat_history, system_message, max_tokens, temperature, top_p, response_style):
    # --- Translation handling ---
    if is_translation_request(message):
        try:
            translated = translator(message, src_lang="auto", tgt_lang="en")[0]["translation_text"]
            chat_history.append({"role": "user", "content": message})
            chat_history.append({"role": "assistant", "content": translated})
            yield "", chat_history
            return
        except Exception as e:
            chat_history.append({"role": "user", "content": message})
            chat_history.append({"role": "assistant", "content": f"⚠️ Translation failed: {str(e)}"})
            yield "", chat_history
            return

    # --- Apply response style ---
    if response_style == "No explanation":
        style_prompt = " Only provide the direct answer with no explanation."
    elif response_style == "Short explanation":
        style_prompt = " Provide a concise answer with a one-sentence explanation."
    else:  # Detailed explanation
        style_prompt = " Provide a thorough and detailed answer with reasoning and examples."

    # --- Prepare messages ---
    # Only keep the last N messages to prevent repetition
    truncated_history = chat_history[-MAX_HISTORY_MESSAGES:]
    messages = [{"role": "system", "content": system_message + style_prompt}] + truncated_history
    messages.append({"role": "user", "content": message})

    # Append user and placeholder for assistant
    chat_history.append({"role": "user", "content": message})
    chat_history.append({"role": "assistant", "content": ""})

    response = ""

    # --- Stream response ---
    for msg in client.chat_completion(
        messages,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        token = msg.choices[0].delta.content or ""
        response += token
        chat_history[-1]["content"] = response
        yield "", chat_history

    # Clear input box after streaming
    yield "", chat_history


# ---------------- UI ----------------
with gr.Blocks(theme=gr.themes.Soft(primary_hue="violet", secondary_hue="pink")) as demo:
    gr.Markdown("# 🤖 Zephyr-7B Chat + 🌍 Translator")

    chatbot = gr.Chatbot(type="messages", height=500, show_copy_button=True, label="Chat Assistant")

    with gr.Row():
        msg = gr.Textbox(label="💬 Your Message", placeholder="Type here…", scale=6)
        send_btn = gr.Button("🚀 Send", variant="primary", scale=1)
        clear_btn = gr.Button("🧹 Clear Chat", scale=1)

    with gr.Accordion("⚙️ Advanced Settings", open=False):
        system_prompt = gr.Textbox(label="System Prompt", value=SYSTEM_PROMPT_DEFAULT, lines=3)
        response_style = gr.Dropdown(
            ["No explanation", "Short explanation", "Detailed explanation"],
            value="No explanation",
            label="Response Style"
        )
        temperature = gr.Slider(0.1, 1.5, value=TEMP_DEFAULT, step=0.1, label="Temperature")
        top_p = gr.Slider(0.1, 1.0, value=TOP_P_DEFAULT, step=0.05, label="Top-p")
        max_tokens = gr.Slider(128, 2048, value=MAX_NEW_TOKENS_DEFAULT, step=16, label="Max new tokens")

    # --- Events ---
    send_btn.click(
        stream_response,
        [msg, chatbot, system_prompt, max_tokens, temperature, top_p, response_style],
        [msg, chatbot]
    )
    msg.submit(
        stream_response,
        [msg, chatbot, system_prompt, max_tokens, temperature, top_p, response_style],
        [msg, chatbot]
    )
    clear_btn.click(lambda: [], None, chatbot, queue=False)

    gr.Markdown("---")
    gr.Markdown("🔗 Built with ❤️ using [Zephyr-7B](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta) & [M2M100](https://huggingface.co/facebook/m2m100_418M).")

if __name__ == "__main__":
    demo.launch()