app.py

import gradio as gr
import datasets
import faiss

from transformers import pipeline


sample_text = """Europejscy astronomowie odkryli planetę
pozasłoneczną pochodzącą spoza naszej galaktyki, czyli 
[unused0] Drogi Mlecznej [unused1]. Obserwacji dokonali
2,2-metrowym teleskopem MPG/ESO."""


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


def load_index(index_data: str = "clarin-knext/entity-linking-index"):
    ds = datasets.load_dataset(index_data)['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/entity-linking-encoder"):
    pipe = pipeline("feature-extraction", model=model_name)
    return pipe


model = load_model()
index = load_index()


def predict(query: str = sample_text, top_k: int=3):
    index_data, faiss_index = index
    # takes only the [CLS] embedding (for now)
    query = model(query, return_tensors = "pt")[0][0].numpy().reshape(1, -1)  
    
    scores, indices = faiss_index.search(query, top_k)
    scores, indices = scores.tolist(), indices.tolist()
    
    results = [
        (index_data[result[0]], result[1])
        for output in zip(indices, scores)
        for result in zip(*output)
    ]
    
    return str(results)


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