app.py

import warnings
warnings.simplefilter(action='ignore', category=FutureWarning)

import PyPDF2
import gradio as gr
from langchain.prompts import PromptTemplate
from langchain.chains.summarize import load_summarize_chain
from huggingface_hub import login
from pathlib import Path
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import os

huggingface_token = os.getenv('HUGGINGFACE_TOKEN')

# Realizar el inicio de sesión de Hugging Face solo si el token está disponible
if huggingface_token:
    login(token=huggingface_token)

# Configuración del modelo de resumen
llm = HuggingFaceEndpoint(
    repo_id="mistralai/Mistral-7B-Instruct-v0.3",
    task="text-generation",
    max_new_tokens=4096,
    temperature=0.5,
    do_sample=False,
)
llm_engine_hf = ChatHuggingFace(llm=llm)

# Configuración del modelo de clasificación
tokenizer = AutoTokenizer.from_pretrained("mrm8488/legal-longformer-base-8192-spanish")
model = AutoModelForSequenceClassification.from_pretrained("mrm8488/legal-longformer-base-8192-spanish")

id2label = {0: "multas", 1: "politicas_de_privacidad", 2: "contratos", 3: "denuncias", 4: "otros"}

def read_pdf(file_path):
    pdf_reader = PyPDF2.PdfReader(file_path)
    text = ""
    for page in range(len(pdf_reader.pages)):
        text += pdf_reader.pages[page].extract_text()
    return text

def summarize(file):
    # Leer el contenido del archivo subido
    file_path = file.name
    if file_path.endswith('.pdf'):
        text = read_pdf(file_path)
    else:
        with open(file_path, 'r', encoding='utf-8') as f:
            text = f.read()
            
    template = '''
    Please carefully read the following document:
<document>
{TEXT}
</document>
After reading through the document, identify the key points and main ideas covered in the text. Organize these key points into a concise bulleted list that summarizes the essential information from the document. The summary should have a maximum of 10 bullet points.
Your goal is to be comprehensive in capturing the core content of the document, while also being concise in how you express each summary point. Omit minor details and focus on the central themes and important facts.
    '''
    
    prompt = PromptTemplate(
        template=template,
        input_variables=['TEXT']    
    )
    
    formatted_prompt = prompt.format(TEXT=text)
    output_summary = llm_engine_hf.invoke(formatted_prompt)
    
    return output_summary.content

def classify_text(text):
    inputs = tokenizer(text, return_tensors="pt", max_length=4096, truncation=True, padding="max_length")
    model.eval()
    with torch.no_grad():
        outputs = model(**inputs)
    logits = outputs.logits
    predicted_class_id = logits.argmax(dim=-1).item()
    predicted_label = id2label[predicted_class_id]
    return predicted_label

def translate(file, target_language):
    # Leer el contenido del archivo subido
    file_path = file.name
    if file_path.endswith('.pdf'):
        text = read_pdf(file_path)
    else:
        with open(file_path, 'r', encoding='utf-8') as f:
            text = f.read()

    template = '''
    Please translate the following document to {LANGUAGE}:
<document>
{TEXT}
</document>
Ensure that the translation is accurate and preserves the original meaning of the document.
    '''
    
    prompt = PromptTemplate(
        template=template,
        input_variables=['TEXT', 'LANGUAGE']    
    )
    
    formatted_prompt = prompt.format(TEXT=text, LANGUAGE=target_language)
    translated_text = llm_engine_hf.invoke(formatted_prompt)
    
    return translated_text

def process_file(file, action, target_language=None):
    if action == "Resumen":
        return summarize(file)
    elif action == "Clasificar":
        file_path = file.name
        if file_path.endswith('.pdf'):
            text = read_pdf(file_path)
        else:
            with open(file_path, 'r', encoding='utf-8') as f:
                text = f.read()
        return classify_text(text)
    elif action == "Traducir":
        return translate(file, target_language)
    else:
        return "Acción no válida"

def download_text(output_text, filename='output.txt'):
    if output_text:
        file_path = Path(filename)
        with open(file_path, 'w', encoding='utf-8') as f:
            f.write(output_text)
        return file_path
    else:
        return None

def create_download_file(output_text, filename='output.txt'):
    file_path = download_text(output_text, filename)
    return str(file_path) if file_path else None

# Crear la interfaz de Gradio
with gr.Blocks() as demo:
    gr.Markdown("## Document Processor")

    with gr.Row():
        with gr.Column():
            file = gr.File(label="Subir un archivo")
            action = gr.Radio(label="Seleccione una acción", choices=["Resumen", "Clasificar", "Traducir"])
            target_language = gr.Dropdown(label="Seleccionar idioma de traducción", choices=["en", "fr", "de"], visible=False)
        
        with gr.Column():
            output_text = gr.Textbox(label="Resultado", lines=20)

    def update_language_dropdown(action):
        if action == "Traducir":
            return gr.update(visible=True)
        else:
            return gr.update(visible=False)
    
    action.change(update_language_dropdown, inputs=action, outputs=target_language)
    
    submit_button = gr.Button("Procesar")
    submit_button.click(process_file, inputs=[file, action, target_language], outputs=output_text)

    def generate_file():
        summary_text = output_text.value
        filename = 'translation.txt' if action.value == 'Traducir' else 'summary.txt'
        file_path = download_text(summary_text, filename)
        return file_path

    download_button = gr.Button("Descargar Resultado")
    download_button.click(
        fn=generate_file,
        inputs=[output_text],
        outputs=gr.File()
    )

# Ejecutar la aplicación Gradio
demo.launch(share=True)