app.py

import os
import spacy
import gradio as gr
from sentence_transformers import SentenceTransformer
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
import zipfile


zip_path = "en_core_web_sm-3.0.0.zip"  # Carica il file ZIP nella cartella del progetto
extraction_dir = "./extracted_models"  # Scegli una sottocartella per l'estrazione
test_dir = "./extracted_models/en_core_web_sm-3.0.0"  # Cartella dopo l'estrazione



# Verifica se la cartella esiste già
if not os.path.exists(test_dir):
    # Se la cartella non esiste, decomprimi il file ZIP
    with zipfile.ZipFile(zip_path, 'r') as zip_ref:
        zip_ref.extractall(extraction_dir)
    print(f"Modello estratto correttamente nella cartella {extraction_dir}")

# Percorso della cartella estratta
model_path = os.path.join(extraction_dir, "en_core_web_sm-3.0.0")  # Assicurati che sia corretto

    # Carica il modello
nlp = spacy.load(model_path)


# Carica il modello SentenceTransformer
#model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2', device='cpu')
#model = SentenceTransformer('sentence-transformers/msmarco-distilbert-base-v4', device='cpu')
#model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2', device='cpu')
model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-mpnet-base-v2', device='cpu')                             
#model = SentenceTransformer('sentence-transformers/all-distilroberta-v1', device='cpu')



# Preprocessamento manuale (carica il manuale da un file o base di dati)
with open('testo.txt', 'r', encoding='utf-8') as file:
    text = file.read()

# Tokenizza il testo in frasi usando SpaCy
doc = nlp(text)
sentences = [sent.text for sent in doc.sents]  # Estrarre frasi dal testo

# Crea gli embedding per il manuale
embeddings = model.encode(sentences, batch_size=8, show_progress_bar=True)

# Funzione per ottenere le frasi più rilevanti
def find_relevant_sentences(query):
    query_embedding = model.encode([query])
    similarities = cosine_similarity(query_embedding, embeddings).flatten()

    # Filtra i risultati in base alla similitudine
    threshold = 0.2
    filtered_results = [(idx, sim) for idx, sim in enumerate(similarities) if sim >= threshold]

    # Ordina i risultati per similitudine
    filtered_results.sort(key=lambda x: x[1], reverse=True)

    # Ottieni le frasi più rilevanti
    top_n = 5
    relevant_sentences = [sentences[idx] for idx, _ in filtered_results[:top_n]]
        
    doc = nlp(" ".join(relevant_sentences))
 
    grouped_results = [sent.text for sent in doc.sents]
    # Pulizia
    cleaned_results = [text.replace("\n", " ") for text in grouped_results]  # Rimuove gli a capo
    final_output = " ".join(cleaned_results)  # Combina tutte le frasi in un unico testo

    
    return final_output

examples = [
    ["irresponsible use of the machine?"],
    ["If I have a problem how can I get help? "],
    ["precautions when using the cutting machine"],
    ["How do I change the knife of the cutting machine?"],
    
]

# Interfaccia Gradio
iface = gr.Interface(
    fn=find_relevant_sentences, 
    inputs=gr.Textbox(label="Insert your query"), 
    outputs=gr.Textbox(label="Relevant sentences"),
    examples=examples, 
    title="Manual Querying System",
    description="Enter a question about the machine, and this tool will find the most relevant sentences from the manual."
)

# Avvia l'app Gradio
iface.launch()