similarity.py

import torch
from sentence_transformers.util import cos_sim
from embeddings import faq_embeddings, data, model
from preprocess import clean_text
import logging

logging.basicConfig(filename="low_score_questions.log", level=logging.INFO, encoding="utf-8")
def get_best_answer(user_question):
    user_question = clean_text(user_question)
    user_embedding = model.encode(
        [user_question],
        convert_to_tensor=True
    )
    similarities = cos_sim(user_embedding, faq_embeddings)
    best_index = torch.argmax(similarities).item()
    best_score = similarities[0][best_index].item()

    adaptive_threshold = 0.55 if len(data) < 500 else 0.5
    if best_score < adaptive_threshold :
        logging.info(f"Unknown question: {user_question}")
        return None
    return data["answer"][best_index]