import streamlit as st import pandas as pd import torch from transformers import pipeline import datetime from rapidfuzz import process, fuzz # Load the CSV file df = pd.read_csv("anomalies.csv", quotechar='"') # Convert 'real' column to standard float format and then to strings df['real'] = df['real'].apply(lambda x: f"{x:.2f}") # Fill NaN values and convert all columns to strings df = df.fillna('').astype(str) # Function to filter the DataFrame using RapidFuzz def filter_dataframe(df, date_str, group_keyword, threshold=80): # Apply fuzzy matching on the 'ds' (date) and 'Group' columns date_matches = process.extract(date_str, df['ds'], scorer=fuzz.token_sort_ratio, limit=None) group_matches = process.extract(group_keyword, df['Group'], scorer=fuzz.token_sort_ratio, limit=None) # Get the indices that match both criteria date_indices = {match[2] for match in date_matches if match[1] >= threshold} group_indices = {match[2] for match in group_matches if match[1] >= threshold} common_indices = list(date_indices & group_indices) return df.iloc[common_indices] # Function to generate a response using the TAPAS model def response(user_question, df): a = datetime.datetime.now() # Extract date and group keywords from the user question date_str = "December 2022" # Example; you'd extract this from the user question dynamically group_keyword = "IPVA" # Filter the DataFrame by date and group subset_df = filter_dataframe(df, date_str, group_keyword) # Check if the DataFrame is empty if subset_df.empty: return {"Resposta": "Desculpe, não há dados disponíveis para responder à sua pergunta."} # Initialize the TAPAS model tqa = pipeline(task="table-question-answering", model="google/tapas-large-finetuned-wtq", tokenizer_kwargs={"clean_up_tokenization_spaces": False}) # Debugging information print("Filtered DataFrame shape:", subset_df.shape) print("Filtered DataFrame head:\n", subset_df.head()) print("User question:", user_question) # Query the TAPAS model try: answer = tqa(table=subset_df, query=user_question)['answer'] except ValueError as e: print(f"Error: {e}") answer = "Desculpe, ocorreu um erro ao processar sua pergunta." query_result = { "Resposta": answer } b = datetime.datetime.now() print("Time taken:", b - a) return query_result # Streamlit interface st.markdown("""