app.py

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()

# Sigmoid-based threshold adjustment function
def calculate_threshold(base_relevance, min_threshold=0.02, max_threshold=0.5, k=10):
    base_relevance_tensor = torch.tensor(base_relevance)  # Ensure input is a tensor
    threshold = min_threshold + (max_threshold - min_threshold) * (
        1 / (1 + torch.exp(-k * (base_relevance_tensor - 0.5)))
    )
    return threshold.item()  # Convert tensor back to float for use in other functions
    
# 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()

    # Compute dynamic threshold using sigmoid-based adjustment
    dynamic_threshold = calculate_threshold(base_relevance_score) * 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), "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), "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), 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=0.5, maximum=2.0, value=1.0, step=0.1, label="Threshold Weight")
    ],
    outputs=[
        gr.Textbox(label="Relevance Score"),
        gr.HTML(label="Highlighted Document Paragraph")
    ],
    title="Cross-Encoder Attention Highlighting",
    description="Adjust the attention threshold weight to control token highlighting sensitivity.",
    allow_flagging="never",
    live=True
)

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