app.py

from transformers import pipeline
from torch import Tensor
from transformers import AutoTokenizer, AutoModel
from torch.nn.functional import cosine_similarity
import gradio as gr

def average_pool(last_hidden_states: Tensor, attention_mask: Tensor) -> Tensor:
    last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
    return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]

def get_similarity(sentence1, sentence2):
  input_texts = [sentence1, sentence2]
  # Tokenize and compute embeddings
  batch_dict = tokenizer(input_texts, max_length=512, padding=True, truncation=True, return_tensors="pt")
  outputs = model(**batch_dict)
  embeddings = average_pool(outputs.last_hidden_state, batch_dict["attention_mask"])
  similarity = cosine_similarity(embeddings[0].unsqueeze(0), embeddings[1].unsqueeze(0))
  similarity = round(similarity.item(), 4)
  return similarity

checkpoint = "intfloat/multilingual-e5-large"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModel.from_pretrained(checkpoint)

demo = gr.Blocks(theme='sudeepshouche/minimalist')
with demo:

    gr.Markdown("# MAGIC Sentence Similarity")
    gr.Markdown("### How to use:")
    gr.Markdown("- Enter Passage 1 and Passage 2, then press Submit")
    gr.Markdown("Model: https://huggingface.co/intfloat/multilingual-e5-large  (Multilingual: 94 languages)")

    with gr.Row():

        p_txt1 = gr.Textbox(placeholder="Enter passage 1", label="Passage 1", lines=3, scale=2)
        p_txt2 = gr.Textbox(placeholder="Enter passage 2", label="Passage 2", lines=3, scale=2)
        o_txt = gr.Textbox(placeholder="Similarity score", lines=1, interactive=False, label="Similarity score (0-1)", scale=1)
   
    submit = gr.Button("Submit")

    submit.click(
        get_similarity,
        [p_txt1, p_txt2],
        o_txt
    )

demo.launch()