Spaces:

ArchCoder
/

basic_app

Sleeping

Update app.py

d6e4129 verified about 1 year ago

7.01 kB

import gradio as gr
from faster_whisper import WhisperModel
from llama_cpp import Llama
from duckduckgo_search import DDGS
import time

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

print("Loading LLM...")
llm = Llama.from_pretrained(
    repo_id="Qwen/Qwen2.5-0.5B-Instruct-GGUF",
    filename="qwen2.5-0.5b-instruct-q4_k_m.gguf",
    n_ctx=2048,
    n_threads=4,
    verbose=False
)

# Initialize DuckDuckGo Search (no API key needed!)
ddgs = DDGS(timeout=3)

def search_web(query, max_results=3):
    """Perform web search using DuckDuckGo (FREE & UNLIMITED)"""
    try:
        # Use text search for fast results
        results = ddgs.text(
            keywords=query,
            region='wt-wt',  # Worldwide results
            safesearch='moderate',
            timelimit='m',  # Last month for freshness
            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 for current info
    search_start = time.time()
    search_results = search_web(question, max_results=2)  # Reduced to 2 for speed
    search_time = time.time() - search_start
    
    # Step 3: Generate answer with LLM
    llm_start = time.time()
    prompt = f"""Answer the question briefly using the context below.
Context:
{search_results}
Question: {question}
Answer:"""
    
    try:
        response = llm(
            prompt,
            max_tokens=120,  # Reduced for faster generation
            temperature=0.2,  # Lower for faster, more focused responses
            top_p=0.85,
            stop=["Question:", "\n\n\n"],
            echo=False
        )
        
        answer = response['choices'][0]['text'].strip()
    except Exception as e:
        answer = f"❌ LLM error: {str(e)}"
    
    llm_time = time.time() - llm_start
    total_time = time.time() - start_time
    
    # Color code timing (green if under 3s, yellow if close, red if over)
    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
with gr.Blocks(title="Fast Q&A - FREE Unlimited Search", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # ⚡ Ultra-Fast Political Q&A System
    Ask questions via audio or text. **FREE unlimited web search** with DuckDuckGo!
    
    **Features:** Whisper-tiny + Qwen2.5-0.5B + DuckDuckGo (No API Key!)
    """)
    
    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?"],
                ["What is the latest news about AI?"]
            ],
            inputs=text_input
        )
    
    with gr.Accordion("📡 API Usage via curl", open=False):
        gr.Markdown("""
        ### Text Query (Simplest):
        ```
        curl -X POST https://archcoder-basic-app.hf.space/call/text_query \\
          -H "Content-Type: application/json" \\
          -d '{"data": ["Who is the current US president?"]}'
        ```
        
        ### Audio Query:
        ```
        # Upload audio
        curl -F "files=@audio.mp3" https://archcoder-basic-app.hf.space/upload
        
        # Query (replace path from upload response)
        curl -X POST https://archcoder-basic-app.hf.space/call/audio_query \\
          -H "Content-Type: application/json" \\
          -d '{"data": [{"path": "/tmp/gradio/YOUR_FILE.mp3"}]}'
        ```
        """)
    
    gr.Markdown("""
    ---
    ### 🎯 System Specs
    - **Search:** DuckDuckGo (FREE, unlimited, no API key!)
    - **Transcription:** Whisper-tiny (optimized for speed)
    - **LLM:** Qwen2.5-0.5B Q4 (fast factual answers)
    - **Target:** Sub-3s total response time
    
    🟢 = Under 3s | 🟡 = 3-3.5s | 🔴 = Over 3.5s
    """)

if __name__ == "__main__":
    demo.queue(max_size=5)  # Limit queue for consistent performance
    demo.launch()

	import gradio as gr
	from faster_whisper import WhisperModel
	from llama_cpp import Llama
	from duckduckgo_search import DDGS
	import time

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

	print("Loading LLM...")
	llm = Llama.from_pretrained(
	repo_id="Qwen/Qwen2.5-0.5B-Instruct-GGUF",
	filename="qwen2.5-0.5b-instruct-q4_k_m.gguf",
	n_ctx=2048,
	n_threads=4,
	verbose=False
	)

	# Initialize DuckDuckGo Search (no API key needed!)
	ddgs = DDGS(timeout=3)

	def search_web(query, max_results=3):
	"""Perform web search using DuckDuckGo (FREE & UNLIMITED)"""
	try:
	# Use text search for fast results
	results = ddgs.text(
	keywords=query,
	region='wt-wt', # Worldwide results
	safesearch='moderate',
	timelimit='m', # Last month for freshness
	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 for current info
	search_start = time.time()
	search_results = search_web(question, max_results=2) # Reduced to 2 for speed
	search_time = time.time() - search_start

	# Step 3: Generate answer with LLM
	llm_start = time.time()
	prompt = f"""Answer the question briefly using the context below.

	Context:
	{search_results}

	Question: {question}

	Answer:"""

	try:
	response = llm(
	prompt,
	max_tokens=120, # Reduced for faster generation
	temperature=0.2, # Lower for faster, more focused responses
	top_p=0.85,
	stop=["Question:", "\n\n\n"],
	echo=False
	)

	answer = response['choices'][0]['text'].strip()
	except Exception as e:
	answer = f"❌ LLM error: {str(e)}"

	llm_time = time.time() - llm_start
	total_time = time.time() - start_time

	# Color code timing (green if under 3s, yellow if close, red if over)
	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
	with gr.Blocks(title="Fast Q&A - FREE Unlimited Search", theme=gr.themes.Soft()) as demo:
	gr.Markdown("""
	# ⚡ Ultra-Fast Political Q&A System
	Ask questions via audio or text. FREE unlimited web search with DuckDuckGo!

	Features: Whisper-tiny + Qwen2.5-0.5B + DuckDuckGo (No API Key!)
	""")

	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?"],
	["What is the latest news about AI?"]
	],
	inputs=text_input
	)

	with gr.Accordion("📡 API Usage via curl", open=False):
	gr.Markdown("""
	### Text Query (Simplest):
	```
	curl -X POST https://archcoder-basic-app.hf.space/call/text_query \\
	-H "Content-Type: application/json" \\
	-d '{"data": ["Who is the current US president?"]}'
	```

	### Audio Query:
	```
	# Upload audio
	curl -F "files=@audio.mp3" https://archcoder-basic-app.hf.space/upload

	# Query (replace path from upload response)
	curl -X POST https://archcoder-basic-app.hf.space/call/audio_query \\
	-H "Content-Type: application/json" \\
	-d '{"data": [{"path": "/tmp/gradio/YOUR_FILE.mp3"}]}'
	```
	""")

	gr.Markdown("""
	---
	### 🎯 System Specs
	- Search: DuckDuckGo (FREE, unlimited, no API key!)
	- Transcription: Whisper-tiny (optimized for speed)
	- LLM: Qwen2.5-0.5B Q4 (fast factual answers)
	- Target: Sub-3s total response time

	🟢 = Under 3s \| 🟡 = 3-3.5s \| 🔴 = Over 3.5s
	""")

	if __name__ == "__main__":
	demo.queue(max_size=5) # Limit queue for consistent performance
	demo.launch()