app.py

import os
import duckdb
import gradio as gr
from dotenv import load_dotenv
from httpx import Client
from huggingface_hub import HfApi
#from llama_cpp import Llama
import pandas as pd
#from transformers import pipeline

load_dotenv()

HF_TOKEN = os.getenv("HF_TOKEN")
assert HF_TOKEN is not None, "You need to set HF_TOKEN in your environment variables"


BASE_DATASETS_SERVER_URL = "https://datasets-server.huggingface.co"
API_URL = "https://m82etjwvhoptr3t5.us-east-1.aws.endpoints.huggingface.cloud"
headers = {
	"Accept" : "application/json",
	"Authorization": f"Bearer {HF_TOKEN}",
	"Content-Type": "application/json" 
}

client = Client(headers=headers)
api = HfApi(token=HF_TOKEN)

# First approach: Use llama.cpp
#llama = Llama(model_path="DuckDB-NSQL-7B-v0.1-q8_0.gguf", n_ctx=2048)
#def query_local_model(text):
#    pred = llama(text, temperature=0.1, max_tokens=500)
#    return pred["choices"][0]["text"]


# Second approach: Use transformers -> Took too much time
#pipe = pipeline("text-generation", model="motherduckdb/DuckDB-NSQL-7B-v0.1")
#def query_local_model_transformers(text):
#    pred = pipe(text, max_length=1000)
#    return pred[0]["generated_text"]


def get_first_parquet(dataset: str):
    resp = client.get(f"{BASE_DATASETS_SERVER_URL}/parquet?dataset={dataset}")
    return resp.json()["parquet_files"][0]


def query_remote_model(text):
    payload = {
        "inputs": text,
        "parameters": {}
    }
    response = client.post(API_URL, headers=headers, json=payload)
    pred = response.json()
    return pred[0]["generated_text"]


def text2sql(dataset_name, query_input):
    print(f"start text2sql for {dataset_name}")
    try:
        first_parquet = get_first_parquet(dataset_name)
    except Exception as error:
        return {
            query_output: "",
            df:pd.DataFrame([{"error": f"❌ Could not get dataset schema. {error=}"}])
        }

    first_parquet_url = first_parquet["url"]
    print(first_parquet_url)
    con = duckdb.connect()
    con.execute("INSTALL 'httpfs'; LOAD httpfs;")
    # could get from Parquet instead?
    con.execute(f"CREATE TABLE data as SELECT * FROM '{first_parquet_url}' LIMIT 1;")
    result = con.sql("SELECT sql FROM duckdb_tables() where table_name ='data';").df()
    ddl_create = result.iloc[0,0]
    
    text = f"""### Instruction:
    Your task is to generate valid duckdb SQL to answer the following question.

    ### Input:
    Here is the database schema that the SQL query will run on:
    {ddl_create}
    
    ### Question:
    {query_input}

    ### Response (use duckdb shorthand if possible) replace table name with {first_parquet_url} in the generated sql query:
    """
    try:
        sql_output =  query_remote_model(text)
    except Exception as error:
        return {
            query_output: "",
            df:pd.DataFrame([{"error": f"❌ Unable to get the SQL query based on the text. {error=}"}])
        }
   
    try:
        query_result = con.sql(sql_output).df()
    except Exception as error:
        query_result = pd.DataFrame([{"error": f"❌ Could not execute SQL query {error=}"}])
    finally:
        con.close()
    return {
        query_output:sql_output,
        df:query_result
    }


with gr.Blocks() as demo:
    gr.Markdown("# Generate SQL queries based on a given text for your dataset")
    gr.Markdown("This space showcase how to generate a SQL query from a text and get the result.")
    gr.Markdown("Tech stack: duckdb and DuckDB-NSQL-7B model")
    dataset_name = gr.Textbox("sksayril/medicine-info", label="Dataset Name")
    query_input = gr.Textbox("How many rows there are?", label="Ask something about your data")
    btn = gr.Button("Generate SQL")
    query_output = gr.Textbox(label="Output SQL", interactive= False)    
    df = gr.DataFrame(datatype="markdown")
    btn.click(text2sql, inputs=[dataset_name, query_input], outputs=[query_output,df])
demo.launch(debug=True)