app.py

import time, os, multiprocessing
from minivectordb.embedding_model import EmbeddingModel
from minivectordb.vector_database import VectorDatabase
from text_util_en_pt.cleaner import structurize_text, detect_language, Language
from webtextcrawler.webtextcrawler import extract_text_from_url
from duckduckgo_search import DDGS
import google.generativeai as genai
import gradio as gr

gemini_key = os.environ.get("GEMINI_KEY")
genai.configure(api_key=gemini_key)
gemini = genai.GenerativeModel('gemini-pro')
model = EmbeddingModel(use_quantized_onnx_model=True)

def fetch_links(query, max_results=10):
    with DDGS() as ddgs:
        return [r['href'] for r in ddgs.text(keywords=query, max_results=max_results)]

def fetch_texts(links):
    with multiprocessing.Pool(10) as pool:
        texts = pool.map(extract_text_from_url, links)
    return '\n'.join([t for t in texts if t])

def index_and_search(query, text):
    start = time.time()
    query_embedding = model.extract_embeddings(query)

    # Indexing
    vector_db = VectorDatabase()
    sentences = [ s['sentence'] for s in structurize_text(text)]

    for idx, sentence in enumerate(sentences):
        sentence_embedding = model.extract_embeddings(sentence)
        vector_db.store_embedding(idx + 1, sentence_embedding, {'sentence': sentence})
    
    embedding_time = time.time() - start

    # Retrieval
    start = time.time()
    search_results = vector_db.find_most_similar(query_embedding, k = 30)
    retrieval_time = time.time() - start
    return '\n'.join([s['sentence'] for s in search_results[2]]), embedding_time, retrieval_time

def generate_search_terms(message, lang):
    if lang == Language.ptbr:
        prompt = f"A partir do texto a seguir, gere alguns termos de pesquisa: \"{message}\"\nSua resposta deve ser apenas o termo de busca mais adequado, e nada mais."
    else:
        prompt = f"From the following text, generate some search terms: \"{message}\"\nYour answer should be just the most appropriate search term, and nothing else."

    response = gemini.generate_content(prompt)
    return response.text

async def predict(message, history):
    full_response = ""

    query_language = detect_language(message)

    start = time.time()
    full_response += "Generating search terms...\n"
    yield full_response
    search_query = generate_search_terms(message, query_language)
    search_terms_time = time.time() - start

    full_response += f"Search terms: \"{search_query}\"\n"
    yield full_response
    full_response += f"Search terms took: {search_terms_time:.4f} seconds\n"
    yield full_response

    start = time.time()
    full_response += "\nSearching the web...\n"
    yield full_response
    links = fetch_links(search_query)
    websearch_time = time.time() - start

    full_response += f"Web search took: {websearch_time:.4f} seconds\n"
    yield full_response

    full_response += f"Links visited:\n"
    yield full_response
    for link in links:
        full_response += f"{link}\n"
        yield full_response
    
    full_response += "\nExtracting text from web pages...\n"
    yield full_response
    start = time.time()
    text = fetch_texts(links)
    webcrawl_time = time.time() - start

    full_response += f"Text extraction took: {webcrawl_time:.4f} seconds\n"

    full_response += "\nIndexing in vector database and building prompt...\n"
    yield full_response

    context, embedding_time, retrieval_time = index_and_search(message, text)

    if query_language == Language.ptbr:
        prompt = f"Contexto:\n{context}\n\nResponda: \"{message}\"\n(Você pode utilizar o contexto para responder)\n(Sua resposta deve ser completa, detalhada e bem estruturada)"
    else:
        prompt = f"Context:\n{context}\n\nAnswer: \"{message}\"\n(You can use the context to answer)\n(Your answer should be complete, detailed and well-structured)"

    full_response += f"Embedding time: {embedding_time:.4f} seconds\n"
    full_response += f"Retrieval from VectorDB time: {retrieval_time:.4f} seconds\n"
    yield full_response

    full_response += "\nGenerating response...\n"
    yield full_response

    full_response += "\nResponse: "
    streaming_response = gemini.generate_content(prompt, stream=True)
    for sr in streaming_response:
        full_response += sr.text
        yield full_response

gr.ChatInterface(
    predict,
    title="Web Search with LLM",
    description="Ask any question, and I will try to answer it using web search",
    retry_btn=None,
    undo_btn=None,
    examples=[
        'When did the first human land on the moon?',
        'Liquid vs solid vs gas?',
        'What is the capital of France?',
        'Why does Brazil has a high tax rate?'
    ]
).launch()