app/dataclean_hf.py

import json
from io import StringIO
import csv
import pandas as pd
import numpy as np
import streamlit as st

from sqlalchemy import create_engine
# from yallmf.utils import run_with_timeout
import openai

from tqdm import tqdm


openai.api_key = st.secrets["OPENAI_API_KEY"]

def clean_data(
    input_product_names: pd.Series, 
    input_brands: pd.Series,
    input_product_categories: pd.Series,
    category_taxonomy: dict):

    output_cols = ['brand', 'product_category', 'sub_product_category', 'strain_name']
    ncols = len(output_cols)
    
    p1 = f'''
    I am going to provide a data set of marijuana products and their metadata. Using the information I provide, I want you to provide me with the following information about the products.

    - Brand (brand)
    - product category (product_category)
    - sub product category (sub_product_category)
    - strain name (strain_name)

    The following JSON shows all the acceptable Product Categories and their Sub Product Categories. Strictly adhere to the below mapping for valid product_category to sub_product_category relationships:

    {json.dumps(category_taxonomy)}
    
    Additional requirements: 

    - The input data set in CSV format, with commas as field delimiter and newline as row delimiter.
    - Do not automatically assume that the information in the data set I provide is accurate.
    - Leave the 'sub_product_category' field blank unless there's a clear and direct match with one of the categories provided in the list.If there is no explicit information to confidently assign a sub_product_category, default to leaving it blank.
    - Strain names are only applicable for the following product categories: concentrate, preroll, vape, flower
    - Look for clues in the product name to determine what brand/ product category/ sub product category/ and strain name the product should fall under. For Vape products, consider the words before 'Cartridge' or 'Cart' in the product name as potential strain names.
    - Every row of the Output CSV must have EXACTLY {ncols} columns.
    - When a field is left empty (e.g., 'sub_product_category' or 'strain_name'), simply leave it empty without placing an additional comma. Each row in the output CSV should always have only three commas separating the four fields regardless of whether some fields are empty. For instance, if 'sub_product_category' and 'strain_name' are empty, a row would look like this: "brand,product_category,,"
    - DO NOT EXPLAIN YOURSELF, ONLY RETURN A CSV WITH THESE COLUMNS: {', '.join(output_cols)}

    Input data set in CSV format:

    '''
    df = pd.DataFrame({'input__product_name':input_product_names,
                    'input__brand':input_brands,
                    'input__product_category':input_product_categories}).reset_index(drop=True)
    # remove commas from all strings
    df2 = df.copy()
    for col in df2.columns:
        df2[col] = df2[col].str.replace(',', '')
    
    # send to LLM
    p2 = df2.to_csv(index=False, quoting=csv.QUOTE_ALL)
    messages = [{'role':'system','content':'You are a helpful assistant. Return a properly-formatted CSV with the correct number of columns.'},
                {'role':'user', 'content':p1+p2+'\n\nOutput CSV with header row:\n\n'}
                ]
    comp = openai.ChatCompletion.create(
        model='gpt-4',
        messages=messages,
        max_tokens=2000,
        timeout=300,
        temperature=0.2
        )
    res = comp['choices'][0]['message']['content']

    # remove rows with wrong number of columns
    keeprows = []
    for i,s in enumerate(res.split('\n')):
        if i==0:
            keeprows.append(s)
            continue
        _ncols = len(s.split(','))
        if _ncols!=ncols:
            print(f'Got {_ncols} columns, skipping row {i-1} ({s})')
            df = df.drop(i-1)
        else:
            keeprows.append(s)
    df = df.reset_index(drop=True)

    resdf = pd.read_csv(StringIO('\n'.join(keeprows)))

    assert len(df)==len(resdf), 'Result CSV did not match input CSV in length'
    df = pd.concat([df.reset_index(drop=True),resdf.reset_index(drop=True)],axis=1)
    # check category/subcategory
    dropidxs=[]
    for idx, row in df.iterrows():
        drop = False
        if pd.isna(row['product_category']) and not pd.isna(row['sub_product_category']):
            drop=True
            print('product_category is null while sub_product_category is not null, dropping')
        if not pd.isna(row['product_category']):
            if row['product_category'] not in category_taxonomy.keys():
                print(f'category "{row["product_category"]}" not in taxonomy, dropping row')
                drop =True
            elif not pd.isna(row['sub_product_category']):
                if row['sub_product_category'] not in category_taxonomy[row['product_category']]:
                    print(f'subcategory "{row["sub_product_category"]}" not valid for category {row["product_category"]}, dropping row')
                    drop =True
        if drop:
            dropidxs.append(idx)
    df = df.drop(dropidxs)

    return df

def get_key(df):
    return df['input__product_name'] + df['input__brand'] + df['input__product_category']

def main(upload_df: pd.DataFrame,
        output_df: pd.DataFrame=None,
        chunksize: int = 30,
        ):
    category_taxonomy = {
        "Wellness": ["Mushroom Caps", "CBD Tincture/Caps/etc", "Promo/ Sample", "Capsule", "Liquid Flower", ""],
        "Concentrate": ["Diamonds", "Shatter", "Sugar", "Promo/ Sample", "Badder", "Diamonds and Sauce", "Rosin", "Cookies Dough", "Flan", "Cookie Dough", ""],
        "Preroll": ["Cubano", "Joint", "Promo/ Sample", "Blunt", "Infused Joint", "Packwoods Blunt", "Infused Blunt", "Napalm", ""],
        "Vape": ["Terp Sauce", "Gpen 0.5", "Cured Resin", "Solventless Rosin", "510", "Dry Flower Series", "Natural Terp Series", "Promo/ Sample", "Dart Pod 0.5", "Raw Garden", "Live Flower Series", "Rosin", "Disposable", ""],
        "Edible": ["Cookies", "Gummies", "Mint", "Promo/ Sample", "Beverage", "Chocolate", ""],
        "Grow Products": ["Promo/ Sample", ""],
        "Flower": ["Promo/ Sample", "Bud", ""],
        "Accessory": ["Promo/ Sample", ""]
    }

    # join together and get the diff
    upload_df = pd.read_csv(upload_df)
    upload_df['key'] = get_key(upload_df)
    upload_df=upload_df.set_index('key')
    if output_df is None:
        rundf = upload_df
        outlen = 0
    else:
        output_df['key'] = get_key(output_df)
        output_df=output_df.set_index('key')
        rundf = upload_df.loc[~upload_df.index.isin(output_df.index)]
        outlen = len(output_df)


    # st.write(f'Input size: {len(upload_df)}, Output size: {outlen}, Still to process: {len(rundf)}')

    for _, chunk in tqdm(rundf.groupby(np.arange(len(rundf)) // chunksize)):
        result = clean_data(chunk['input__product_name'], chunk['input__brand'], chunk['input__product_category'], category_taxonomy)
        result['key'] = get_key(result)
        result = result.set_index('key')
        output_df = result if output_df is None else pd.concat([output_df, result])
    
    return output_df