import gradio as gr
import torch
from transformers import BertTokenizer, BertForMaskedLM

# Load the fine-tuned BERT model
model_name = "fine_tuned_bert_model"
tokenizer = BertTokenizer.from_pretrained(model_name)
model = BertForMaskedLM.from_pretrained(model_name)
model.to("cuda" if torch.cuda.is_available() else "cpu")

# Function to answer questions using the fine-tuned model
def answer_question(context, question):
    # Preprocess the context and question
    context_tokens = tokenizer(context, truncation=True, padding="max_length", max_length=128, return_tensors="pt")
    question_tokens = tokenizer(question, truncation=True, padding="max_length", max_length=16, return_tensors="pt")

    # Move tensors to device
    context_tokens = context_tokens.to(model.device)
    question_tokens = question_tokens.to(model.device)

    with torch.no_grad():
        # Generate masked LM predictions for each token in the question
        outputs = model(**question_tokens)
        predictions = torch.argmax(outputs.logits, dim=-1)

    # Replace masked tokens in the question with predicted tokens
    answer_tokens = []
    for i in range(len(question_tokens["input_ids"][0])):
        if question_tokens["input_ids"][0][i] == tokenizer.mask_token_id:
            answer_tokens.append(predictions[0][i].item())
        else:
            answer_tokens.append(question_tokens["input_ids"][0][i].item())

    # Decode tokens and remove special tokens
    answer = tokenizer.decode(answer_tokens, skip_special_tokens=True)

    # Return the answer
    return answer

# Define example questions
examples = [
    ["Where did the Enron scandal occur?", "The Enron scandal occurred in [MASK]."],
    ["What was the outcome of the Enron scandal?", "The outcome of the Enron scandal was [MASK]."],
    ["When did Enron file for bankruptcy?", "Enron filed for bankruptcy in [MASK]."],
    ["How did Enron's stock price change during the scandal?", "During the Enron scandal, Enron's stock price [MASK]."]
]

# Gradio interface with examples
iface = gr.Interface(
    fn=answer_question,
    inputs=["text", "text"],
    outputs="text",
    title="Enron Email Analysis",
    description="Ask questions about the Enron email dataset using a fine-tuned BERT model.",
    examples=examples
)

# Launch the Gradio interface
iface.launch(share=True)