app.py

import gradio as gr
from huggingface_hub import InferenceClient
import json
import uuid
from PIL import Image
from bs4 import BeautifulSoup
import requests
import random
from transformers import LlavaProcessor, LlavaForConditionalGeneration, TextIteratorStreamer
from threading import Thread
import re
import time 
import torch
import cv2

model_id = "llava-hf/llava-interleave-qwen-0.5b-hf"

processor = LlavaProcessor.from_pretrained(model_id)

model = LlavaForConditionalGeneration.from_pretrained(model_id, low_cpu_mem_usage=True)
model.to("cpu")


def sample_frames(video_file) :
    try:
        video = cv2.VideoCapture(video_file)
        total_frames = int(video.get(cv2.CAP_PROP_FRAME_COUNT))
        num_frames = 12
        interval = total_frames // num_frames
        frames = []
        for i in range(total_frames):
            ret, frame = video.read()
            pil_img = Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
            if not ret:
                continue
            if i % interval == 0:
                frames.append(pil_img)
        video.release()
        return frames
    except:
        frames=[]
        return frames

def extract_text_from_webpage(html_content):
    soup = BeautifulSoup(html_content, 'html.parser')
    for tag in soup(["script", "style", "header", "footer"]):
        tag.extract()
    return soup.get_text(strip=True)

def search(query):
    term = query
    start = 0
    all_results = []
    max_chars_per_page = 8000
    with requests.Session() as session:
        resp = session.get(
            url="https://www.google.com/search",
            headers={"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:109.0) Gecko/20100101 Firefox/111.0"},
            params={"q": term, "num": 3, "udm": 14},
            timeout=5,
            verify=None,
        )
        resp.raise_for_status()
        soup = BeautifulSoup(resp.text, "html.parser")
        result_block = soup.find_all("div", attrs={"class": "g"})
        for result in result_block:
            link = result.find("a", href=True)
            link = link["href"]
            try:
                webpage = session.get(link, headers={"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64; rv:109.0) Gecko/20100101 Firefox/111.0"}, timeout=5, verify=False)
                webpage.raise_for_status()
                visible_text = extract_text_from_webpage(webpage.text)
                if len(visible_text) > max_chars_per_page:
                    visible_text = visible_text[:max_chars_per_page]
                all_results.append({"link": link, "text": visible_text})
            except requests.exceptions.RequestException:
                all_results.append({"link": link, "text": None})
    return all_results

# Initialize inference clients for different models
client_gemma = InferenceClient("google/gemma-1.1-7b-it")
client_mixtral = InferenceClient("NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO")
client_llama = InferenceClient("meta-llama/Meta-Llama-3-8B-Instruct")

# Define the main chat function
def respond(message, history):
    func_caller = []
    vqa = ""

    # Handle image processing
    if message["files"]:
        image = user_prompt["files"][-1]    
        txt = user_prompt["text"]
        img = user_prompt["files"]
        
        video_extensions = ("avi", "mp4", "mov", "mkv", "flv", "wmv", "mjpeg", "wav", "gif", "webm", "m4v", "3gp")
        image_extensions = Image.registered_extensions()
        image_extensions = tuple([ex for ex, f in image_extensions.items()])
            
        if image.endswith(video_extensions):
            image = sample_frames(image)
            image_tokens = "<image>" * int(len(image))
            prompt = f"<|im_start|>user {image_tokens}\n{user_prompt}<|im_end|><|im_start|>assistant"
              
        elif image.endswith(image_extensions):
            image = Image.open(image).convert("RGB")
            prompt = f"<|im_start|>user <image>\n{user_prompt}<|im_end|><|im_start|>assistant"
    
        print(len(image))
    
        inputs = processor(prompt, image, return_tensors="pt")
        streamer = TextIteratorStreamer(processor, skip_prompt=True, **{"skip_special_tokens": True})
        generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=1024)
        generated_text = ""
    
        thread = Thread(target=model.generate, kwargs=generation_kwargs)
        thread.start()
    
        buffer = ""
        for new_text in streamer:
            buffer += new_text
            yield buffer
        

    # Define function metadata for user interface
    functions_metadata = [
        {"type": "function", "function": {"name": "web_search", "description": "Search query on google", "parameters": {"type": "object", "properties": {"query": {"type": "string", "description": "web search query"}}, "required": ["query"]}}},
        {"type": "function", "function": {"name": "general_query", "description": "Reply general query of USER", "parameters": {"type": "object", "properties": {"prompt": {"type": "string", "description": "A detailed prompt"}}, "required": ["prompt"]}}},
        {"type": "function", "function": {"name": "image_generation", "description": "Generate image for user", "parameters": {"type": "object", "properties": {"query": {"type": "string", "description": "image generation prompt"}, "number_of_image": {"type": "integer", "description": "number of images to generate"}}, "required": ["query"]}}},
        {"type": "function", "function": {"name": "image_qna", "description": "Answer question asked by user related to image", "parameters": {"type": "object", "properties": {"query": {"type": "string", "description": "Question by user"}}, "required": ["query"]}}},
    ]

    message_text = message["text"]
    func_caller.append({"role": "user", "content": f'[SYSTEM]You are a helpful assistant. You have access to the following functions: \n {str(functions_metadata)}\n\nTo use these functions respond with:\n<functioncall> {{ "name": "function_name", "arguments": {{ "arg_1": "value_1", "arg_1": "value_1", ... }} }}  </functioncall>  [USER] {message} {vqa}'})

    response = client_gemma.chat_completion(func_caller, max_tokens=150)
    response = str(response)
    try:
        response = response[int(response.find("{")):int(response.index("</"))]
    except:
        print("A error occured")
    response = response.replace("\\n", "")
    response = response.replace("\\'", "'")
    response = response.replace('\\"', '"')
    print(f"\n{response}")

    func_caller.append({"role": "assistant", "content": f"<functioncall>{response}</functioncall>"})

    try:
        json_data = json.loads(str(response))
        if json_data["name"] == "web_search":
            query = json_data["arguments"]["query"]
            gr.Info("Searching Web")
            web_results = search(query)
            gr.Info("Extracting relevant Info")
            web2 = ' '.join([f"Link: {res['link']}\nText: {res['text']}\n\n" for res in web_results])
            messages = f"<|im_start|>system\nYou are OpenGPT 4o mini a helpful assistant made by KingNish. You are provided with WEB results from which you can find informations to answer users query in Structured and More better way. You do not say Unnecesarry things Only say thing which is important and relevant. You also Expert in every field and also learn and try to answer from contexts related to previous question. Try your best to give best response possible to user. You also try to show emotions using Emojis and reply like human, use short forms, friendly tone and emotions.<|im_end|>"
            for msg in history:
                messages += f"\n<|im_start|>user\n{str(msg[0])}<|im_end|>"
                messages += f"\n<|im_start|>assistant\n{str(msg[1])}<|im_end|>"
            messages+=f"\n<|im_start|>user\n{message_text} {vqa}<|im_end|>\n<|im_start|>web_result\n{web2}<|im_end|>\n<|im_start|>assistant\n"
            stream = client_mixtral.text_generation(messages, max_new_tokens=2000, do_sample=True, stream=True, details=True, return_full_text=False)
            output = ""
            for response in stream:
                if not response.token.text == "<|im_end|>":
                    output += response.token.text
                    yield output
        elif json_data["name"] == "image_generation":
            query = json_data["arguments"]["query"]
            gr.Info("Generating Image, Please wait...")
            seed = random.randint(1, 99999)
            query = query.replace(" ", "%20")
            image = f"![](https://image.pollinations.ai/prompt/{query}?seed={seed})"
            yield image
            gr.Info("We are going to Update Our Image Generation Engine to more powerful ones in Next Update. ThankYou")
        elif json_data["name"] == "image_qna":
            messages = f"<|start_header_id|>system\nYou are OpenGPT 4o mini a helpful assistant made by KingNish. You are provide with both images and captions and Your task is to answer of user with help of caption provided. Answer in human style and show emotions.<|end_header_id|>"
            for msg in history:
                messages += f"\n<|start_header_id|>user\n{str(msg[0])}<|end_header_id|>"
                messages += f"\n<|start_header_id|>assistant\n{str(msg[1])}<|end_header_id|>"
            messages+=f"\n<|start_header_id|>user\n{message_text} {vqa}<|end_header_id|>\n<|start_header_id|>assistant\n"
            stream = client_llama.text_generation(messages, max_new_tokens=2000, do_sample=True, stream=True, details=True, return_full_text=False)
            output = ""
            for response in stream:
                if not response.token.text == "<|eot_id|>":
                    output += response.token.text
                    yield output
        else:
            messages = f"<|start_header_id|>system\nYou are OpenGPT 4o mini a helpful assistant made by KingNish. You answers users query like human friend. You are also Expert in every field and also learn and try to answer from contexts related to previous question. Try your best to give best response possible to user. You also try to show emotions using Emojis and reply like human, use short forms, friendly tone and emotions.<|end_header_id|>"
            for msg in history:
                messages += f"\n<|start_header_id|>user\n{str(msg[0])}<|end_header_id|>"
                messages += f"\n<|start_header_id|>assistant\n{str(msg[1])}<|end_header_id|>"
            messages+=f"\n<|start_header_id|>user\n{message_text} {vqa}<|end_header_id|>\n<|start_header_id|>assistant\n"
            stream = client_llama.text_generation(messages, max_new_tokens=2000, do_sample=True, stream=True, details=True, return_full_text=False)
            output = ""
            for response in stream:
                if not response.token.text == "<|eot_id|>":
                    output += response.token.text
                    yield output
    except:
        messages = f"<|start_header_id|>system\nYou are OpenGPT 4o mini a helpful assistant made by KingNish. You answers users query like human friend. You are also Expert in every field and also learn and try to answer from contexts related to previous question. Try your best to give best response possible to user. You also try to show emotions using Emojis and reply like human, use short forms, friendly tone and emotions.<|end_header_id|>"
        for msg in history:
            messages += f"\n<|start_header_id|>user\n{str(msg[0])}<|end_header_id|>"
            messages += f"\n<|start_header_id|>assistant\n{str(msg[1])}<|end_header_id|>"
        messages+=f"\n<|start_header_id|>user\n{message_text} {vqa}<|end_header_id|>\n<|start_header_id|>assistant\n"
        stream = client_llama.text_generation(messages, max_new_tokens=2000, do_sample=True, stream=True, details=True, return_full_text=False)
        output = ""
        for response in stream:
            if not response.token.text == "<|eot_id|>":
                output += response.token.text
                yield output

# Create the Gradio interface
demo = gr.ChatInterface(
    fn=respond,
    chatbot=gr.Chatbot(show_copy_button=True, likeable=True, layout="panel"),
    title="OpenGPT 4o mini",
    textbox=gr.MultimodalTextbox(),
    multimodal=True,
    concurrency_limit=20,
    examples=[
        {"text": "Hy, who are you?",},
        {"text": "What's the current price of Bitcoin",},
        {"text": "Create A Beautiful image of Effiel Tower at Night",},
        {"text": "Write me a Python function to calculate the first 10 digits of the fibonacci sequence.",},
        {"text": "What's the colour of both of Car in given image", "files": ["./car1.png", "./car2.png"]},
    ],
    cache_examples=False,
)
demo.launch()