app.py

import gradio as gr
import datasets
import faiss
import os

from transformers import pipeline


auth_token = os.environ.get("CLARIN_KNEXT")


sample_text = (
    "Wydarzenia te miały miejsce na początku mojej przyjaźni z Holmesem, "
    "kiedy, jeszcze jako [unused0] kawalerowie [unused1], mieszkaliśmy razem przy Baker Street."
)


textbox = gr.Textbox(
    label="Type your query here.",
    value=sample_text, lines=10
)


def load_index(index_data: str = "clarin-knext/wsd-linking-index"):
    ds = datasets.load_dataset(index_data, use_auth_token=auth_token)['train']
    index_data = {
        idx: (e_id, e_text) for idx, (e_id, e_text) in
        enumerate(zip(ds['entities'], ds['texts']))
    }
    faiss_index = faiss.read_index("./encoder.faissindex", faiss.IO_FLAG_MMAP)
    return index_data, faiss_index


def load_model(model_name: str = "clarin-knext/wsd-encoder"):
    model = pipeline("feature-extraction", model=model_name, use_auth_token=auth_token)
    return model


model = load_model()
index = load_index()


def predict(text: str = sample_text, top_k: int=3):
    index_data, faiss_index = index
    # takes only the [CLS] embedding (for now)
    query = model(text, return_tensors='pt')[0][0].numpy().reshape(1, -1)

    scores, indices = faiss_index.search(query, top_k)
    scores, indices = scores.tolist(), indices.tolist()

    results = "\n".join([
        f"{index_data[result[0]]}: {result[1]}"
        for output in zip(indices, scores)
        for result in zip(*output)
    ])

    return results


demo = gr.Interface(fn=predict, inputs=textbox, outputs="text").launch()