app.py

import gradio as gr
import duckdb
from annoy import AnnoyIndex
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("sentence-transformers/LaBSE")

annoy_index = AnnoyIndex(768, "angular")
annoy_index.load("definitions.ann")

conn = duckdb.connect("sonajaht.db")


def search_query(query, top_k=10):
    query_vector = model.encode(query)

    similar_item_ids, distances = annoy_index.get_nns_by_vector(
        query_vector, top_k, include_distances=True
    )

    id_list = ", ".join(map(str, similar_item_ids))
    sql_query = f"""
    SELECT w.value AS word, d.value AS definition 
    FROM definitions d
    JOIN words w ON d.word_id = w.word_id
    WHERE d.entry_id IN ({id_list})
    ORDER BY CASE d.entry_id 
    {' '.join([f'WHEN {_id} THEN {i}' for i, _id in enumerate(similar_item_ids)])}
    END
    """

    results = conn.execute(sql_query).fetchdf()

    results["relevance_score"] = [1 - d for d in distances]

    return results


# Example queries in natural language
examples = [
    "väga vana mees",
    "очень старый дед",
    "un très vieil homme"
]


def handle_example(example):
    return example, search_query(example)


with gr.Blocks() as demo:
    gr.Markdown("# Sõnajaht Demo")

    query_input = gr.Textbox(label="Sisestage teie otsingupäring")
    search_button = gr.Button("Otsi")

    with gr.Row():
        example_buttons = [gr.Button(example) for example in examples]

    results_output = gr.Dataframe(label="Otsingutulemused")

    search_button.click(search_query, inputs=query_input, outputs=results_output)

    for button in example_buttons:
        button.click(
            handle_example,
            inputs=gr.State(button.value),
            outputs=[query_input, results_output],
        )


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