app.py

import streamlit as st
import pandas as pd
import zipfile
import os
from sentence_transformers import SentenceTransformer, util
from transformers import pipeline

# Constants
ZIP_FILE = "xnli-multilingual-nli-dataset.zip"
CSV_FILE = "en_test.csv"
EXTRACT_FOLDER = "extracted_data"

# Load and extract ZIP
@st.cache_data
def extract_and_load():
    if not os.path.exists(EXTRACT_FOLDER):
        with zipfile.ZipFile(ZIP_FILE, "r") as zip_ref:
            zip_ref.extractall(EXTRACT_FOLDER)
    csv_path = os.path.join(EXTRACT_FOLDER, CSV_FILE)
    df = pd.read_csv(csv_path).dropna().sample(500)
    return df[['premise', 'hypothesis', 'label']]

df = extract_and_load()

# Load models
nli_model = pipeline("text-classification", model="joeddav/xlm-roberta-large-xnli")
embedder = SentenceTransformer("sentence-transformers/distiluse-base-multilingual-cased-v2")

# UI
st.title("🌐 Multilingual RAG-style NLI Explorer")
st.markdown("Enter a sentence in **any language**, and the app will find a related statement from the dataset and infer their relationship.")

user_input = st.text_input("Enter your **hypothesis** (your own sentence):")

if user_input:
    with st.spinner("Finding most relevant premise..."):
        premise_embeddings = embedder.encode(df['premise'].tolist(), convert_to_tensor=True)
        user_embedding = embedder.encode(user_input, convert_to_tensor=True)

        top_hit = util.semantic_search(user_embedding, premise_embeddings, top_k=1)[0][0]
        match_idx = top_hit['corpus_id']
        selected_premise = df.iloc[match_idx]['premise']

        st.subheader("🔍 Most Relevant Premise:")
        st.write(selected_premise)

        # Run NLI classification
        full_input = f"{selected_premise} </s> {user_input}"
        result = nli_model(full_input)[0]

        st.subheader("🧠 Predicted Relationship:")
        st.write(f"**{result['label']}** (confidence: {result['score']:.2f})")