import torch
import pandas as pd
import gradio as gr
from datasets import load_dataset
from sentence_transformers import SentenceTransformer, util, models


q_encoder = SentenceTransformer(modules=[
    models.Transformer(model_name_or_path="checkpoints/q_encoder", max_seq_length=512),
    models.Pooling(word_embedding_dimension=768, pooling_mode='cls'),
])
doc_embeddings = torch.load('checkpoints/doc_embeddings.pt', map_location=torch.device('cpu'))
docs = pd.DataFrame(load_dataset("maastrichtlawtech/bsard", data_files="articles_fr.csv")['train'])

def search(question):
    q_emb = q_encoder.encode(question, convert_to_tensor=True)
    hits = util.semantic_search(q_emb, doc_embeddings, top_k=100, score_function=util.cos_sim)[0]
    return {docs.loc[h['corpus_id'], 'article'] + '\n\n' + f"- Art. {docs.loc[h['corpus_id'], 'article_no']}, {docs.loc[h['corpus_id'], 'code']}" for h in hits[:5]}

title = "Belgian Legislation Search"
description = "A biencoder model was trained to retrieve relevant statutory articles to legal issues. Ask it a question in French!"
article = """
The model will return the most semantically relevant laws from a corpus of 22,633 statutory articles collected from 32 Belgian codes.
"""
examples = [
    "Qu'est-ce que je risque si je viole le secret professionnel ?",
    "Mon employeur peut-il me licencier alors que je suis malade ?",
    "Mon voisin fait beaucoup de bruit, que faire ?",
]
gr.Interface(fn=search, inputs=['text'], outputs=['textbox']*5, allow_flagging="never", title=title, description=description, article=article, examples=examples).launch()