app.py

import requests
import gradio as gr
from numpy import dot
from numpy.linalg import norm
from os import environ

num_of_sentences = 2

endpoint = environ.get("ENDPOINT", "")
api_key = environ.get("API_KEY", "")


def cos_sim(a, b):
    return dot(a, b) / (norm(a) * norm(b))


def get_embeddings(model, text):
    # make POST request to the endpoint
    response = requests.post(
        f"{endpoint}/{model}",
        json={"text": text},
        headers={"Authorization": "Bearer " + api_key},
    )

    return response.json()["vector"]


def calculate_similarities(model, text, *sentences):
    # get embeddings for the input text
    text_embedding = get_embeddings(model, text)

    # get embeddings for the input sentences
    sentences_embeddings = [get_embeddings(model, sentence) for sentence in sentences]

    # calculate cosine similarity between the input text and the input sentences
    similarities = {}

    # to normalize cosine similarity to be between 0 and 1
    minx = -1
    maxx = 1

    for sentence, sentence_embedding in zip(sentences, sentences_embeddings):
        sim = cos_sim(text_embedding, sentence_embedding)
        normalized_sim = (sim - minx) / (maxx - minx)
        similarities[sentence] = normalized_sim

    return similarities


demo = gr.Blocks()

with demo:
    with gr.Row():
        with gr.Column():
            model = gr.inputs.Radio(
                ["roberta", "ada"], default="roberta", label="Model"
            )

            text = gr.Textbox(lines=3, label="Input Text")

            inp_sentences = [
                gr.Textbox(lines=3, label="Sentence " + str(i + 1))
                for i in range(num_of_sentences)
            ]
            btn = gr.Button(text="Submit")

        with gr.Column():
            output = gr.Label(label="Output", show_label=False)

            # submit btn
            btn.click(
                calculate_similarities,
                inputs=[model, text, *inp_sentences],
                outputs=[output],
            )

    gr.Examples(
        examples=[
            ["roberta", "This is happy person", "هذا شخص سعيد", "هذه قطة سعيدة"],
            ["ada", "This is happy person", "هذا شخص سعيد", "هذه قطة سعيدة"],
            ["roberta", "هذا شخص سعيد", "هذه قطة سعيدة", "This is happy person"],
            ["ada", "هذا شخص سعيد", "هذه قطة سعيدة", "This is happy person"],
            ["roberta", "car", "camry", "toyota"],
            ["ada", "camry", "toy", "toyota"],
            ["roberta", "ihpone for sale", "iphone for sale", "camry for sale"],
            ["ada", "ihpone for sale", "iphone for sale", "camry for sale"],
        ],
        inputs=[model, text, *inp_sentences],
        outputs=output,
        fn=calculate_similarities,
        # cache_examples=True,
    )

if __name__ == "__main__":
    demo.launch()