app.py

import duckdb
import gradio as gr
import polars as pl
from datasets import load_dataset
from gradio_huggingfacehub_search import HuggingfaceHubSearch
from model2vec import StaticModel

global ds
global df

# Load a model from the HuggingFace hub (in this case the potion-base-8M model)
model_name = "minishlab/M2V_multilingual_output"
model = StaticModel.from_pretrained(model_name)


def get_iframe(hub_repo_id):
    if not hub_repo_id:
        raise ValueError("Hub repo id is required")
    url = f"https://huggingface.co/datasets/{hub_repo_id}/embed/viewer"
    iframe = f"""
    <iframe
  src="{url}"
  frameborder="0"
  width="100%"
  height="600px"
></iframe>
"""
    return iframe


def load_dataset_from_hub(hub_repo_id):
    global ds
    ds = load_dataset(hub_repo_id)


def get_columns(split: str):
    global ds
    ds_split = ds[split]
    return gr.Dropdown(
        choices=ds_split.column_names,
        value=ds_split.column_names[0],
        label="Select a column",
    )


def get_splits():
    global ds
    splits = list(ds.keys())
    return gr.Dropdown(choices=splits, value=splits[0], label="Select a split")


def vectorize_dataset(split: str, column: str):
    global df
    global ds
    df = ds[split].to_polars()
    embeddings = model.encode(df[column])
    df = df.with_columns(pl.Series(embeddings).alias("embeddings"))


def run_query(query: str):
    global df
    vector = model.encode(query)
    return duckdb.sql(
        query=f"""
        SELECT *
        FROM df
        ORDER BY array_distance(embeddings, {vector.tolist()}::FLOAT[256])
        LIMIT 5
        """
    ).to_df()


with gr.Blocks() as demo:
    gr.HTML(
        """
        <h1>Vector Search any Hugging Face Dataset</h1>
        <p>
            This app allows you to vector search any Hugging Face dataset.
            You can search for the nearest neighbors of a query vector, or
            perform a similarity search on a dataframe.
        </p>
        <p>
            This app uses the <a href="https://huggingface.co/minishlab/M2V_multilingual_output">M2V_multilingual_output</a> model from the Hugging Face Hub.
        </p>
        """
    )
    with gr.Row():
        with gr.Column():
            search_in = HuggingfaceHubSearch(
                label="Search Huggingface Hub",
                placeholder="Search for models on Huggingface",
                search_type="dataset",
                sumbit_on_select=True,
            )
    with gr.Row():
        search_out = gr.HTML(label="Search Results")
    search_in.submit(get_iframe, inputs=search_in, outputs=search_out)

    btn_load_dataset = gr.Button("Load Dataset")

    with gr.Row(variant="panel"):
        split_dropdown = gr.Dropdown(label="Select a split")
        column_dropdown = gr.Dropdown(label="Select a column")
    btn_vectorize_dataset = gr.Button("Vectorize")

    btn_load_dataset.click(
        load_dataset_from_hub, inputs=search_in, show_progress=True
    ).then(fn=get_splits, outputs=split_dropdown).then(
        fn=get_columns, inputs=split_dropdown, outputs=column_dropdown
    )
    split_dropdown.change(
        fn=get_columns, inputs=split_dropdown, outputs=column_dropdown
    )

    btn_vectorize_dataset.click(
        fn=vectorize_dataset,
        inputs=[split_dropdown, column_dropdown],
        show_progress=True,
    )

    with gr.Row(variant="panel"):
        query_input = gr.Textbox(label="Query")

    btn_run = gr.Button("Run")
    results_output = gr.Dataframe(label="Results")

    btn_run.click(fn=run_query, inputs=query_input, outputs=results_output)
demo.launch()