app.py

# Inspired by https://huggingface.co/spaces/asoria/duckdb-parquet-demo

import gradio as gr
import duckdb
import pandas as pd
import requests

DATASETS_SERVER_ENDPOINT = "https://datasets-server.huggingface.co"
PARQUET_REVISION="refs/convert/parquet"

EXAMPLE_DATASET_NAME = "LLMs/Alpaca-ShareGPT"


def get_parquet_urls(dataset: str) -> list[str]:
        splits = requests.get(f"{DATASETS_SERVER_ENDPOINT}/splits?dataset={dataset}", timeout=60).json().get("splits")
        split = splits[0]
        response = requests.get(f"{DATASETS_SERVER_ENDPOINT}/parquet?dataset={dataset}&config={split['config']}", timeout=60)
        if response.status_code != 200:
            raise Exception(response)
        
        response = response.json()
        parquet_files = response["parquet_files"]
        urls = [content["url"] for content in parquet_files if content["split"] == split["split"]]
        if len(urls) == 0:
             raise Exception("No parquet files found for dataset")
        return urls

def run_command(query: str) -> pd.DataFrame:
    try:
        result = duckdb.execute("SELECT fts_main_data.match_bm25(id, ?) AS score, id, instruction, input, output   FROM data   WHERE score IS NOT NULL   ORDER BY score;", [query])
        print("Ok")
    except Exception as error:
        print(f"Error: {str(error)}")
        return pd.DataFrame({"Error": [f"❌ {str(error)}"]})
    print(result)
    return result.df()

def import_data():
    # Import data + index
    parquet_url = get_parquet_urls(EXAMPLE_DATASET_NAME)[0]
    print("parquet_url", parquet_url)
    duckdb.sql("CREATE SEQUENCE serial START 1;")
    # We need a sequence id column for Full text search
    # I'm very rusty in SQL so it's very possible there are simpler ways.

    duckdb.sql(f"CREATE TABLE data AS SELECT nextval('serial') AS id, * FROM '{parquet_url}';")
    duckdb.sql("PRAGMA create_fts_index('data', 'id', '*');")

    duckdb.sql("DESCRIBE SELECT * FROM data").show()
    print("foo foo")



with gr.Blocks() as demo:
    gr.Markdown(" ## Full-text search using DuckDB on top of datasets-server Parquet files 🐤")
    gr.CheckboxGroup(label="Dataset", choices=["LLMs/Alpaca-ShareGPT"], value="LLMs/Alpaca-ShareGPT", info="Dataset to query"),
    query = gr.Textbox(label="query", placeholder="Full-text search...")
    run_button = gr.Button("Run")
    gr.Markdown("### Result")
    cached_responses_table = gr.DataFrame()
    run_button.click(run_command, inputs=[query], outputs=cached_responses_table)


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