import gradio as gr
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "cross-encoder/ms-marco-MiniLM-L-12-v2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
model.eval()

# Function to compute relevance score and dynamically adjust threshold
def get_relevance_score_and_excerpt(query, paragraph, threshold_weight):
    if not query.strip() or not paragraph.strip():
        return "Please provide both a query and a document paragraph.", "", ""

    # Tokenize the input
    inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True)
    
    with torch.no_grad():
        output = model(**inputs, output_attentions=True)

    # Extract logits and calculate base relevance score
    logit = output.logits.squeeze().item()
    base_relevance_score = torch.sigmoid(torch.tensor(logit)).item()

    # Dynamically adjust the attention threshold based on user weight
    dynamic_threshold = max(0.02, threshold_weight)

    # Extract attention scores (last layer)
    attention = output.attentions[-1]
    attention_scores = attention.mean(dim=1).mean(dim=0)

    query_tokens = tokenizer.tokenize(query)
    paragraph_tokens = tokenizer.tokenize(paragraph)

    query_len = len(query_tokens) + 2  # +2 for special tokens [CLS] and first [SEP]
    para_start_idx = query_len
    para_end_idx = len(inputs["input_ids"][0]) - 1

    if para_end_idx <= para_start_idx:
        return round(base_relevance_score, 4), round(dynamic_threshold, 4), "No relevant tokens extracted."

    para_attention_scores = attention_scores[para_start_idx:para_end_idx, para_start_idx:para_end_idx].mean(dim=0)

    if para_attention_scores.numel() == 0:
        return round(base_relevance_score, 4), round(dynamic_threshold, 4), "No relevant tokens extracted."

    # Get indices of relevant tokens above dynamic threshold
    relevant_indices = (para_attention_scores > dynamic_threshold).nonzero(as_tuple=True)[0].tolist()

    # Reconstruct paragraph with bolded relevant tokens using HTML tags
    highlighted_text = ""
    for idx, token in enumerate(paragraph_tokens):
        if idx in relevant_indices:
            highlighted_text += f"<b>{token}</b> "
        else:
            highlighted_text += f"{token} "

    highlighted_text = tokenizer.convert_tokens_to_string(highlighted_text.split())

    return round(base_relevance_score, 4), round(dynamic_threshold, 4), highlighted_text

# Define Gradio interface with a slider for threshold adjustment
interface = gr.Interface(
    fn=get_relevance_score_and_excerpt,
    inputs=[
        gr.Textbox(label="Query", placeholder="Enter your search query..."),
        gr.Textbox(label="Document Paragraph", placeholder="Enter a paragraph to match..."),
        gr.Slider(minimum=0.02, maximum=0.5, value=0.1, step=0.01, label="Attention Threshold")
    ],
    outputs=[
        gr.Textbox(label="Relevance Score"),
        gr.Textbox(label="Dynamic Threshold"),
        gr.HTML(label="Highlighted Document Paragraph")
    ],
    title="Cross-Encoder Attention Highlighting",
    description="Adjust the attention threshold to control token highlighting sensitivity.",
    allow_flagging="never",
    live=True
)

if __name__ == "__main__":
    interface.launch()