app.py

import gradio as gr
from faster_whisper import WhisperModel
from transformers import AutoTokenizer, AutoModelForCausalLM
from duckduckgo_search import DDGS
import time
import torch

# Initialize models
print("Loading Whisper model...")
whisper_model = WhisperModel("tiny", device="cpu", compute_type="int8")

print("Loading LLM...")
model_name = "Qwen/Qwen2.5-0.5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float32,
    device_map="cpu",
    low_cpu_mem_usage=True
)

# Initialize DuckDuckGo Search
ddgs = DDGS(timeout=3)

def search_web(query, max_results=2):
    """Perform web search using DuckDuckGo (FREE & UNLIMITED)"""
    try:
        results = ddgs.text(
            keywords=query,
            region='wt-wt',
            safesearch='moderate',
            timelimit='m',
            max_results=max_results
        )
        
        context = ""
        for i, result in enumerate(results[:max_results], 1):
            title = result.get('title', '')
            body = result.get('body', '')
            context += f"\n[{i}] {title}\n{body}\n"
        
        return context.strip() if context else "No search results found."
    
    except Exception as e:
        return f"Search failed: {str(e)}"

def process_audio(audio_path, question_text=None):
    """Main pipeline: audio -> text -> search -> answer"""
    start_time = time.time()
    
    # Step 1: Transcribe audio if provided
    if audio_path:
        try:
            segments, _ = whisper_model.transcribe(audio_path, language="en", beam_size=1)
            question = " ".join([seg.text for seg in segments])
        except Exception as e:
            return f"❌ Transcription error: {str(e)}", 0.0
    else:
        question = question_text
    
    if not question or question.strip() == "":
        return "❌ No input provided", 0.0
    
    transcription_time = time.time() - start_time
    
    # Step 2: Web search
    search_start = time.time()
    search_results = search_web(question, max_results=2)
    search_time = time.time() - search_start
    
    # Step 3: Generate answer with LLM
    llm_start = time.time()
    
    messages = [
        {"role": "system", "content": "You are a helpful assistant. Answer questions briefly using the provided context."},
        {"role": "user", "content": f"Context:\n{search_results}\n\nQuestion: {question}\n\nAnswer:"}
    ]
    
    try:
        text = tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True
        )
        
        inputs = tokenizer([text], return_tensors="pt").to("cpu")
        
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=120,
                temperature=0.2,
                do_sample=True,
                top_p=0.85,
                pad_token_id=tokenizer.eos_token_id
            )
        
        response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
        answer = response.strip()
        
    except Exception as e:
        answer = f"❌ LLM error: {str(e)}"
    
    llm_time = time.time() - llm_start
    total_time = time.time() - start_time
    
    time_emoji = "🟢" if total_time < 3.0 else "🟡" if total_time < 3.5 else "🔴"
    timing_info = f"\n\n{time_emoji} **Timing:** Trans={transcription_time:.2f}s | Search={search_time:.2f}s | LLM={llm_time:.2f}s | **Total={total_time:.2f}s**"
    
    return answer + timing_info, total_time

# Create Gradio interface (same as before)
with gr.Blocks(title="Fast Q&A - No Building Required!", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # ⚡ Ultra-Fast Political Q&A System
    **No wheel building** - Fast deployment with transformers!
    
    **Features:** Whisper-tiny + Qwen2.5-0.5B + DuckDuckGo (FREE unlimited search)
    """)
    
    with gr.Tab("🎙️ Audio Input"):
        with gr.Row():
            with gr.Column():
                audio_input = gr.Audio(
                    sources=["microphone", "upload"],
                    type="filepath",
                    label="Record or upload audio"
                )
                audio_submit = gr.Button("🚀 Submit Audio", variant="primary", size="lg")
            
            with gr.Column():
                audio_output = gr.Textbox(label="Answer", lines=8, show_copy_button=True)
                audio_time = gr.Number(label="Response Time (seconds)", precision=2)
        
        audio_submit.click(
            fn=lambda x: process_audio(x, None),
            inputs=[audio_input],
            outputs=[audio_output, audio_time],
            api_name="audio_query"
        )
    
    with gr.Tab("✍️ Text Input"):
        with gr.Row():
            with gr.Column():
                text_input = gr.Textbox(
                    label="Type your question",
                    placeholder="Who is the current US president?",
                    lines=3
                )
                text_submit = gr.Button("🚀 Submit Text", variant="primary", size="lg")
            
            with gr.Column():
                text_output = gr.Textbox(label="Answer", lines=8, show_copy_button=True)
                text_time = gr.Number(label="Response Time (seconds)", precision=2)
        
        text_submit.click(
            fn=lambda x: process_audio(None, x),
            inputs=[text_input],
            outputs=[text_output, text_time],
            api_name="text_query"
        )
        
        gr.Examples(
            examples=[
                ["Who won the 2024 US presidential election?"],
                ["What is the current inflation rate in India?"],
                ["Who is the prime minister of UK?"]
            ],
            inputs=text_input
        )
    
    gr.Markdown("""
    ---
    🟢 = Under 3s | 🟡 = 3-3.5s | 🔴 = Over 3.5s
    """)

if __name__ == "__main__":
    demo.queue(max_size=5)
    demo.launch()