app.py

import os
import requests
import yt_dlp
import whisper
import torch
import gradio as gr
from bs4 import BeautifulSoup
from groq import Groq
from youtube_transcript_api import YouTubeTranscriptApi

# Setup using Hugging Face Secrets
ZENROWS_KEY = os.environ.get("ZENROWS_KEY")
GROQ_KEY = os.environ.get("GROQ_KEY")

client = Groq(api_key=GROQ_KEY)

# Detect Hardware
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"🚀 Running on: {device.upper()}")

# Load Whisper (This will happen when the Hugging Face Space starts)
audio_model = whisper.load_model("base.en").to(device)

cache = {"web_url": "", "web_content": "", "yt_url": "", "yt_content": ""}

def get_website_content(url):
    global cache
    url = url.strip()
    if url == cache["web_url"]: return cache["web_content"]
    
    params = {"apikey": ZENROWS_KEY, "url": url, "js_render": "true", "premium_proxy": "true"}
    try:
        res = requests.get("https://api.zenrows.com/v1/", params=params, timeout=30)
        soup = BeautifulSoup(res.text, "html.parser")
        for junk in soup(["script", "style", "nav", "footer", "header"]): junk.decompose()
        text = " ".join([p.text for p in soup.find_all('p')])
        cache["web_url"], cache["web_content"] = url, text[:12000]
        return cache["web_content"]
    except Exception as e: return f"Error: {str(e)}"

def get_youtube_content(url):
    global cache
    video_id = url.split("v=")[-1].split("&")[0].split("/")[-1].strip()
    if video_id == cache["yt_url"]: return cache["yt_content"]

    try:
        transcript = YouTubeTranscriptApi.get_transcript(video_id)
        content = " ".join([i['text'] for i in transcript])
    except:
        ydl_opts = {
            'format': 'm4a/bestaudio/best',
            'outtmpl': 'temp_audio.%(ext)s',
            'quiet': True,
            'postprocessors': [{'key': 'FFmpegExtractAudio','preferredcodec': 'm4a'}],
        }
        with yt_dlp.YoutubeDL(ydl_opts) as ydl: ydl.download([url])
        result = audio_model.transcribe("temp_audio.m4a", fp16=(device=="cuda"))
        content = result["text"]
        if os.path.exists("temp_audio.m4a"): os.remove("temp_audio.m4a")

    cache["yt_url"], cache["yt_content"] = video_id, content
    return content

def ask_ai(context, question):
    if not context or len(context) < 20: return "❌ Error: No content found."
    completion = client.chat.completions.create(
        model="llama-3.3-70b-versatile",
        messages=[
            {"role": "system", "content": f"Answer using ONLY this text:\n\n{context}"},
            {"role": "user", "content": question}
        ]
    )
    return completion.choices[0].message.content

with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 🚀 Ultimate AI Research Assistant (V2)")
    with gr.Tabs():
        with gr.TabItem("🌐 Website Analysis"):
            url_w = gr.Textbox(label="Website URL")
            q_w = gr.Textbox(label="Question")
            out_w = gr.Textbox(label="AI Response")
            gr.Button("Analyze Site").click(lambda u,q: ask_ai(get_website_content(u), q), [url_w, q_w], out_w)
        with gr.TabItem("🎥 YouTube Analysis"):
            url_y = gr.Textbox(label="YouTube URL")
            q_y = gr.Textbox(label="Question")
            out_y = gr.Textbox(label="AI Response")
            gr.Button("Analyze Video").click(lambda u,q: ask_ai(get_youtube_content(u), q), [url_y, q_y], out_y)

demo.launch()