excel_chat.py

import gradio as gr
from mistralai.client import MistralClient
from mistralai.models.chat_completion import ChatMessage
import os
import pandas as pd
import numpy as np
from groq import Groq
import anthropic
from users_management import update_json, users


def ask_llm(query, input, client_index):
    messages = [
                {
                    "role": "system",
                    "content": f"You are a helpful assistant. Only show your final response to the **User Query**! Do not provide any explanations or details: \n# User Query:\n{query}."
                },
                {
                    "role": "user",
                    "content": f"{input}",
                }
            ]
    systemC = f"You are a helpful assistant. Only show your final response to the **User Query**! Do not provide any explanations or details: \n# User Query:\n{query}."   
    messageC=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": f"{input}"
                }
            ]
        }
    ]

    if client_index == "Groq":
        client = Groq(api_key=os.environ["GROQ_API_KEY"])
        chat_completion = client.chat.completions.create(
            messages=messages,
            model='mixtral-8x7b-32768',
        )
    elif client_index == "Mistral":
        client = MistralClient(api_key=os.environ['MISTRAL_API_KEY'])
        chat_completion = client.chat(
            messages=messages,
            model='mistral-small-latest',
        )
    else:
        client = anthropic.Anthropic(api_key=os.environ['CLAUDE_API_KEY'])
        chat_completion = client.messages.create(
            model="claude-3-sonnet-20240229",
            max_tokens=350,
            temperature=0,
            system=systemC,
            messages=messageC
        ).content[0].text
        return chat_completion

    return chat_completion.choices[0].message.content

def filter_df(df, column_name, keywords):
    if len(keywords)>0:
        if column_name in df.columns:
            contains_keyword = lambda x: any(keyword.lower() in (x.lower() if type(x)==str else '') for keyword in keywords)
            filtered_df = df[df[column_name].apply(contains_keyword)]
        else:
            contains_keyword = lambda row: any(keyword.lower() in (str(cell).lower() if isinstance(cell, str) else '') for keyword in keywords for cell in row)
            filtered_df = df[df.apply(contains_keyword, axis=1)]
    else:
        filtered_df = df
    return filtered_df

def chat_with_mistral(source_cols, dest_col, prompt, excel_file, url, search_col, keywords, client, user):
    update_json(user, prompt, keywords)
    print(f'xlsxfile = {excel_file}')
    df = pd.read_excel(excel_file)
    df[dest_col] = ""
    try:
      file_name = url.split("/")[-2] + ".xlsx"
    except:
      file_name = excel_file
    print(f"Keywords: {keywords}")

    filtred_df = filter_df(df, search_col, keywords)

    for index, row in filtred_df.iterrows():
        concatenated_content = "\n\n".join(f"{column_name}: {str(row[column_name])}" for column_name in source_cols)
        print('test')
        if not concatenated_content == "\n\n".join(f"{column_name}: nan" for column_name in source_cols):
            print('c bon')
            llm_answer = ask_llm(prompt, concatenated_content, client)
            print(f"QUERY:\n{prompt}\nCONTENT:\n{concatenated_content[:200]}...\n\nANSWER:\n{llm_answer}")
            df.at[index, dest_col] = llm_answer

    df.to_excel(file_name, index=False)
    return file_name, df.head(5)


def get_columns(file,progress=gr.Progress()):
    if file is not None:
        df = pd.read_excel(file)
        columns = list(df.columns)
        return gr.update(choices=columns), gr.update(choices=columns), gr.update(choices=columns), gr.update(choices=columns + [""]), gr.update(choices=columns + ['[ALL]']), df.head(5)
    else:
        return gr.update(choices=[]), gr.update(choices=[]), gr.update(choices=[]), gr.update(choices=[]), gr.update(choices=[]), pd.DataFrame()