app.py

import os
import duckdb
import gradio as gr
from dotenv import load_dotenv
from httpx import Client
from huggingface_hub import HfApi
import pandas as pd
from gradio_huggingfacehub_search import HuggingfaceHubSearch
import spaces
from llama_cpp import Llama

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 = "/static-proxy?url=https%3A%2F%2Fdatasets-server.huggingface.co%26quot%3B%3C%2Fspan%3E%3C!----%3E%3C%2Ftd%3E%3C%2Ftr%3E%3Ctr id="L18">
headers = {
	"Accept" : "application/json",
	"Authorization": f"Bearer {HF_TOKEN}",
	"Content-Type": "application/json" 
}
client = Client(headers=headers)
api = HfApi(token=HF_TOKEN)
llama = Llama(
        model_path="DuckDB-NSQL-7B-v0.1-q8_0.gguf",
        n_ctx=2048,
        n_gpu_layers=50
    )

@spaces.GPU
def generate_sql(prompt):
    # pred = pipe(prompt, max_length=1000)
    # return pred[0]["generated_text"]
    pred = llama(prompt, temperature=0.1, max_tokens=1000)
    return pred["choices"][0]["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 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 {
            schema_output: "",
            prompt_output: "",
            query_output: "",
            df:pd.DataFrame([{"error": f"❌ Could not get dataset schema. {error=}"}])
        }

    first_parquet_url = first_parquet["url"]
    print(f"getting schema from {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):
    """
    try:
        sql_output = generate_sql(text)
    except Exception as error:
        return {
            schema_output: ddl_create,
            prompt_output: text,
            query_output: "",
            df:pd.DataFrame([{"error": f"❌ Unable to get the SQL query based on the text. {error=}"}])
        }

    # Should be replaced by the prompt but not working
    sql_output = sql_output.replace("FROM data", f"FROM '{first_parquet_url}'")
    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 {
        schema_output: ddl_create,
        prompt_output: text,
        query_output:sql_output,
        df:query_result
    }


with gr.Blocks() as demo:
    gr.Markdown("# 💫 Generate SQL queries based on a given text for your Hugging Face Dataset 💫")
    dataset_name = HuggingfaceHubSearch(
            label="Hub Dataset ID",
            placeholder="Search for dataset id on Huggingface",
            search_type="dataset",
            value="jamescalam/world-cities-geo",
        )
    # dataset_name = gr.Textbox("jamescalam/world-cities-geo", label="Dataset Name")
    query_input = gr.Textbox("Cities from Albania country", label="Ask something about your data")
    examples = [
                ["Cities from Albania country"],
                ["The continent with the most number of countries"],
                ["Cities that start with 'A'"],
                ["Cities by region"],
            ]
    gr.Examples(examples=examples, inputs=[query_input],outputs=[])
    btn = gr.Button("Generate SQL")
    query_output = gr.Textbox(label="Output SQL", interactive= False)    
    df = gr.DataFrame(datatype="markdown")
    with gr.Accordion("Open for prompt details", open=False):
        #with gr.Column(scale=1, min_width=600):
        schema_output = gr.Textbox(label="Parquet Schema as CREATE DDL", interactive= False)
        prompt_output = gr.Textbox(label="Generated prompt", interactive= False)
    btn.click(text2sql, inputs=[dataset_name, query_input], outputs=[schema_output, prompt_output, query_output,df])
demo.launch(debug=True)