app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForMaskedLM
import torch
import numpy as np
from tqdm.auto import tqdm
import os


# CSS to style the custom share button (for the "Sparse Representation" tab)
css = """
.share-button-container {
    display: flex;
    justify-content: center;
    margin-top: 10px;
    margin-bottom: 20px;
}

.custom-share-button {
    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
    border: none;
    color: white;
    padding: 8px 16px;
    text-align: center;
    text-decoration: none;
    display: inline-block;
    font-size: 14px;
    margin: 4px 2px;
    cursor: pointer;
    border-radius: 6px;
    transition: all 0.3s ease;
}

.custom-share-button:hover {
    transform: translateY(-2px);
    box-shadow: 0 4px 8px rgba(0,0,0,0.2);
}
/* IMPORTANT: This CSS targets Gradio's *default* share button that appears
   when demo.launch(share=True) is used.
   You might want to comment this out if you prefer Gradio's default positioning
   for the main share button (usually in the header/footer) and rely only on your custom one.
*/
.share-button {
    position: fixed !important;
    top: 20px !important;
    right: 20px !important;
    z-index: 1000 !important;
    background: #4CAF50 !important;
    color: white !important;
    border-radius: 8px !important;
    padding: 8px 16px !important;
    font-weight: bold !important;
    box-shadow: 0 2px 10px rgba(0,0,0,0.2) !important;
}

.share-button:hover {
    background: #45a049 !important;
    transform: translateY(-1px) !important;
}
"""


# --- Model Loading ---
tokenizer_splade = None
model_splade = None
tokenizer_splade_lexical = None
model_splade_lexical = None
tokenizer_splade_doc = None
model_splade_doc = None

# Load SPLADE v3 model (original)
try:
    tokenizer_splade = AutoTokenizer.from_pretrained("naver/splade-cocondenser-selfdistil")
    model_splade = AutoModelForMaskedLM.from_pretrained("naver/splade-cocondenser-selfdistil")
    model_splade.eval()
    print("SPLADE-cocondenser-distil model loaded successfully!")
except Exception as e:
    print(f"Error loading SPLADE-cocondenser-distil model: {e}")
    print("Please ensure you have accepted any user access agreements on the Hugging Face Hub page for 'naver/splade-cocondenser-selfdistil'.")

# Load SPLADE v3 Lexical model
try:
    splade_lexical_model_name = "naver/splade-v3-lexical"
    tokenizer_splade_lexical = AutoTokenizer.from_pretrained(splade_lexical_model_name)
    model_splade_lexical = AutoModelForMaskedLM.from_pretrained(splade_lexical_model_name)
    model_splade_lexical.eval()
    print(f"SPLADE-v3-Lexical model '{splade_lexical_model_name}' loaded successfully!")
except Exception as e:
    print(f"Error loading SPLADE-v3-Lexical model: {e}")
    print(f"Please ensure '{splade_lexical_model_name}' is accessible (check Hugging Face Hub for potential agreements).")

# Load SPLADE v3 Doc model - Model loading is still necessary even if its logits aren't used for BoW
try:
    splade_doc_model_name = "naver/splade-v3-doc"
    tokenizer_splade_doc = AutoTokenizer.from_pretrained(splade_doc_model_name)
    model_splade_doc = AutoModelForMaskedLM.from_pretrained(splade_doc_model_name) # Still load the model
    model_splade_doc.eval()
    print(f"SPLADE-v3-Doc model '{splade_doc_model_name}' loaded successfully!")
except Exception as e:
    print(f"Error loading SPLADE-v3-Doc model: {e}")
    print(f"Please ensure '{splade_doc_model_name}' is accessible (check Hugging Face Hub for potential agreements).")


# --- Helper function for lexical mask (now handles batches, but used for single input here) ---
def create_lexical_bow_mask(input_ids_batch, vocab_size, tokenizer):
    """
    Creates a batch of lexical BOW masks.
    input_ids_batch: torch.Tensor of shape (batch_size, sequence_length)
    vocab_size: int, size of the tokenizer vocabulary
    tokenizer: the tokenizer object
    Returns: torch.Tensor of shape (batch_size, vocab_size)
    """
    batch_size = input_ids_batch.shape[0]
    bow_masks = torch.zeros(batch_size, vocab_size, device=input_ids_batch.device)

    for i in range(batch_size):
        input_ids = input_ids_batch[i] # Get input_ids for the current item in the batch
        meaningful_token_ids = []
        for token_id in input_ids.tolist():
            if token_id not in [
                tokenizer.pad_token_id,
                tokenizer.cls_token_id,
                tokenizer.sep_token_id,
                tokenizer.mask_token_id,
                tokenizer.unk_token_id
            ]:
                meaningful_token_ids.append(token_id)

        if meaningful_token_ids:
            # Apply mask to the current row in the batch
            bow_masks[i, list(set(meaningful_token_ids))] = 1

    return bow_masks


# --- Core Representation Functions (Return Formatted Strings - for Explorer Tab) ---
# These functions now return a tuple: (main_representation_str, info_str)
def get_splade_cocondenser_representation(text):
    if tokenizer_splade is None or model_splade is None:
        return "SPLADE-cocondenser-distil model is not loaded. Please check the console for loading errors.", ""

    inputs = tokenizer_splade(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(model_splade.device) for k, v in inputs.items()}

    with torch.no_grad():
        output = model_splade(**inputs)

    if hasattr(output, 'logits'):
        splade_vector = torch.max(
            torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
            dim=1
        )[0].squeeze() # Squeeze is fine here as it's a single input
    else:
        return "Model output structure not as expected for SPLADE-cocondenser-distil. 'logits' not found.", ""

    indices = torch.nonzero(splade_vector).squeeze().cpu().tolist()
    if not isinstance(indices, list):
        indices = [indices] if indices else []

    values = splade_vector[indices].cpu().tolist()
    token_weights = dict(zip(indices, values))

    meaningful_tokens = {}
    for token_id, weight in token_weights.items():
        decoded_token = tokenizer_splade.decode([token_id])
        if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0:
            meaningful_tokens[decoded_token] = weight

    sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[1], reverse=True)

    formatted_output = "MLM Representation:\n\n"
    if not sorted_representation:
        formatted_output += "No significant terms found for this input.\n"
    else:
        # Changed to paragraph style
        terms_list = []
        for term, weight in sorted_representation:
            terms_list.append(f"**{term}**: {weight:.4f}")
        formatted_output += ", ".join(terms_list) + "."

    info_output = f"" # Line 1
    info_output += f"Total non-zero terms in vector: {len(indices)}\n" # Line 2 (and onwards for sparsity)


    return formatted_output, info_output


def get_splade_lexical_representation(text):
    if tokenizer_splade_lexical is None or model_splade_lexical is None:
        return "SPLADE-v3-Lexical model is not loaded. Please check the console for loading errors.", ""

    inputs = tokenizer_splade_lexical(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(model_splade_lexical.device) for k, v in inputs.items()}

    with torch.no_grad():
        output = model_splade_lexical(**inputs)

    if hasattr(output, 'logits'):
        splade_vector = torch.max(
            torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
            dim=1
        )[0].squeeze() # Squeeze is fine here
    else:
        return "Model output structure not as expected for SPLADE-v3-Lexical. 'logits' not found.", ""

    # Always apply lexical mask for this model's specific behavior
    vocab_size = tokenizer_splade_lexical.vocab_size
    # Call with unsqueezed input_ids for single sample processing
    bow_mask = create_lexical_bow_mask(
        inputs['input_ids'], vocab_size, tokenizer_splade_lexical
    ).squeeze() # Squeeze back for single output
    splade_vector = splade_vector * bow_mask

    indices = torch.nonzero(splade_vector).squeeze().cpu().tolist()
    if not isinstance(indices, list):
        indices = [indices] if indices else []

    values = splade_vector[indices].cpu().tolist()
    token_weights = dict(zip(indices, values))

    meaningful_tokens = {}
    for token_id, weight in token_weights.items():
        decoded_token = tokenizer_splade_lexical.decode([token_id])
        if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0:
            meaningful_tokens[decoded_token] = weight

    sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[1], reverse=True)

    formatted_output = "MLP Representation:\n\n"
    if not sorted_representation:
        formatted_output += "No significant terms found for this input.\n"
    else:
        # Changed to paragraph style
        terms_list = []
        for term, weight in sorted_representation:
            terms_list.append(f"**{term}**: {weight:.4f}")
        formatted_output += ", ".join(terms_list) + "."

    info_output = f"" # Line 1
    info_output += f"Total non-zero terms in vector: {len(indices)}\n" # Line 2 (and onwards for sparsity)


    return formatted_output, info_output


def get_splade_doc_representation(text):
    if tokenizer_splade_doc is None: # No longer need model_splade_doc to be loaded for 'logits'
        return "SPLADE-v3-Doc tokenizer is not loaded. Please check the console for loading errors.", ""

    inputs = tokenizer_splade_doc(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(torch.device("cpu")) for k, v in inputs.items()} # Ensure on CPU for direct mask creation

    vocab_size = tokenizer_splade_doc.vocab_size
    # Directly create the binary Bag-of-Words vector using the input_ids
    binary_bow_vector = create_lexical_bow_mask(
        inputs['input_ids'], vocab_size, tokenizer_splade_doc
    ).squeeze() # Squeeze back for single output

    indices = torch.nonzero(binary_bow_vector).squeeze().cpu().tolist()
    if not isinstance(indices, list):
        indices = [indices] if indices else []

    values = [1.0] * len(indices) # All values are 1 for binary representation
    token_weights = dict(zip(indices, values))

    meaningful_tokens = {}
    for token_id, weight in token_weights.items():
        decoded_token = tokenizer_splade_doc.decode([token_id])
        if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0:
            meaningful_tokens[decoded_token] = weight

    sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[0]) # Sort alphabetically for clarity

    formatted_output = "Binary Bag-of-Words Representation:\n\n"
    if not sorted_representation:
        formatted_output += "No significant terms found for this input.\n"
    else:
        # Changed to paragraph style
        terms_list = []
        for term, _ in sorted_representation: # For binary, weight is always 1, so no need to display
            terms_list.append(f"**{term}**")
        formatted_output += ", ".join(terms_list) + "."

    info_output = f"" # Line 1
    info_output += f"Total non-zero terms in vector: {len(indices)}" # Line 2

    return formatted_output, info_output


# --- Unified Prediction Function for the Explorer Tab ---
def predict_representation_explorer(model_choice, text):
    if model_choice == "MLM encoder (SPLADE-cocondenser-distil)":
        return get_splade_cocondenser_representation(text)
    elif model_choice == "MLP encoder (SPLADE-v3-lexical)":
        return get_splade_lexical_representation(text)
    elif model_choice == "Binary Bag-of-Words": # Changed name
        return get_splade_doc_representation(text)
    else:
        return "Please select a model.", "" # Return two empty strings for consistency

# --- Core Representation Functions (Return RAW TENSORS - for Dot Product Tab) ---
# These functions remain unchanged from the previous iteration, as they return the raw tensors.
def get_splade_cocondenser_vector(text):
    if tokenizer_splade is None or model_splade is None:
        return None

    inputs = tokenizer_splade(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(model_splade.device) for k, v in inputs.items()}

    with torch.no_grad():
        output = model_splade(**inputs)

    if hasattr(output, 'logits'):
        splade_vector = torch.max(
            torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
            dim=1
        )[0].squeeze()
        return splade_vector
    return None

def get_splade_lexical_vector(text):
    if tokenizer_splade_lexical is None or model_splade_lexical is None:
        return None

    inputs = tokenizer_splade_lexical(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(model_splade_lexical.device) for k, v in inputs.items()}

    with torch.no_grad():
        output = model_splade_lexical(**inputs)

    if hasattr(output, 'logits'):
        splade_vector = torch.max(
            torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
            dim=1
        )[0].squeeze()

        vocab_size = tokenizer_splade_lexical.vocab_size
        bow_mask = create_lexical_bow_mask(
            inputs['input_ids'], vocab_size, tokenizer_splade_lexical
        ).squeeze()

        splade_vector = splade_vector * bow_mask
        return splade_vector
    return None

def get_splade_doc_vector(text):
    if tokenizer_splade_doc is None: # No longer need model_splade_doc to be loaded for 'logits'
        return None

    inputs = tokenizer_splade_doc(text, return_tensors="pt", padding=True, truncation=True)
    inputs = {k: v.to(torch.device("cpu")) for k, v in inputs.items()} # Ensure on CPU for direct mask creation

    vocab_size = tokenizer_splade_doc.vocab_size
    # Directly create the binary Bag-of-Words vector using the input_ids
    binary_bow_vector = create_lexical_bow_mask(
        inputs['input_ids'], vocab_size, tokenizer_splade_doc
    ).squeeze()
    return binary_bow_vector


# --- Function to get formatted representation from a raw vector and tokenizer ---
# This function remains unchanged as it's a generic formatter for any sparse vector.
def format_sparse_vector_output(splade_vector, tokenizer, is_binary=False):
    if splade_vector is None:
        return "Failed to generate vector.", ""

    indices = torch.nonzero(splade_vector).squeeze().cpu().tolist()
    if not isinstance(indices, list):
        indices = [indices] if indices else []

    if is_binary:
        values = [1.0] * len(indices)
    else:
        values = splade_vector[indices].cpu().tolist()

    token_weights = dict(zip(indices, values))

    meaningful_tokens = {}
    for token_id, weight in token_weights.items():
        decoded_token = tokenizer.decode([token_id])
        if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0:
            meaningful_tokens[decoded_token] = weight

    if is_binary:
        sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[0]) # Sort alphabetically for binary
    else:
        sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[1], reverse=True)

    formatted_output = ""
    if not sorted_representation:
        formatted_output += "No significant terms found.\n"
    else:
        terms_list = []
        for i, (term, weight) in enumerate(sorted_representation):
            if i >= 50:
                terms_list.append(f"...and {len(sorted_representation) - 50} more terms.")
                break
            if is_binary:
                terms_list.append(f"**{term}**")
            else:
                terms_list.append(f"**{term}**: {weight:.4f}")
        formatted_output += ", ".join(terms_list) + "."

    # This is the line that will now always be split into two
    info_output = f"Total non-zero terms: {len(indices)}\n" # Line 1


    return formatted_output, info_output


# --- NEW/MODIFIED: Helper to get the correct vector function, tokenizer, and binary flag ---
def get_model_assets(model_choice_str):
    if model_choice_str == "MLM encoder (SPLADE-cocondenser-distil)":
        return get_splade_cocondenser_vector, tokenizer_splade, False, "MLM encoder (SPLADE-cocondenser-distil)"
    elif model_choice_str == "MLP encoder (SPLADE-v3-lexical)":
        return get_splade_lexical_vector, tokenizer_splade_lexical, False, "MLP encoder (SPLADE-v3-lexical)"
    elif model_choice_str == "Binary Bag-of-Words":
        return get_splade_doc_vector, tokenizer_splade_doc, True, "Binary Bag-of-Words"
    else:
        return None, None, False, "Unknown Model"

# --- MODIFIED: Dot Product Calculation Function for the new tab ---
def calculate_dot_product_and_representations_independent(query_model_choice, doc_model_choice, query_text, doc_text):
    query_vector_fn, query_tokenizer, query_is_binary, query_model_name_display = get_model_assets(query_model_choice)
    doc_vector_fn, doc_tokenizer, doc_is_binary, doc_model_name_display = get_model_assets(doc_model_choice)

    if query_vector_fn is None or doc_vector_fn is None:
        return "Please select valid models for both query and document encoding.", "", ""

    query_vector = query_vector_fn(query_text)
    doc_vector = doc_vector_fn(doc_text)

    if query_vector is None or doc_vector is None:
        return "Failed to generate one or both vectors. Please check model loading and input text.", "", ""

    # Calculate dot product
    # Ensure both vectors are on CPU before dot product to avoid device mismatch issues
    # and to ensure .item() works reliably for conversion to float.
    dot_product = float(torch.dot(query_vector.cpu(), doc_vector.cpu()).item())

    # Format representations - these functions now return two strings (main_output, info_output)
    query_main_rep_str, query_info_str = format_sparse_vector_output(query_vector, query_tokenizer, query_is_binary)
    doc_main_rep_str, doc_info_str = format_sparse_vector_output(doc_vector, doc_tokenizer, doc_is_binary)


    # Combine output into a single string for the Markdown component
    full_output = f"### Dot Product Score: {dot_product:.6f}\n\n"
    full_output += "---\n\n"

    # Query Representation
    full_output += f"Query Representation ({query_model_name_display}):\n\n"
    full_output += query_main_rep_str + "\n\n" + query_info_str # Added an extra newline for better spacing
    full_output += "\n\n---\n\n" # Separator

    # Document Representation
    full_output += f"Document Representation ({doc_model_name_display}):\n\n"
    full_rep_str = doc_main_rep_str + "\n\n" + doc_info_str # Adjusted for consistency
    full_output += full_rep_str

    return full_output


# Global variable to store the share URL once the app is launched
# This will be populated by the demo.launch() call
global_share_url = None

def get_current_share_url():
    """Returns the globally stored share URL."""
    return global_share_url if global_share_url else "Share URL not available yet."

# --- Gradio Interface Setup with Tabs ---
with gr.Blocks(title="SPLADE Demos", css=css) as demo:
    gr.Markdown("# 🌌 Sparse Encoder Playground") # Updated title
    gr.Markdown("Explore different SPLADE models and their sparse representation types, and calculate similarity between query and document representations.") # Updated description

    with gr.Tabs():
        with gr.TabItem("Sparse Representation"):
            gr.Markdown("### Produce a Sparse Representation of an Input Text")
            with gr.Row():
                with gr.Column(scale=1): # Left column for inputs and info
                    model_radio = gr.Radio(
                        [
                            "MLM encoder (SPLADE-cocondenser-distil)",
                            "MLP encoder (SPLADE-v3-lexical)",
                            "Binary Bag-of-Words"
                        ],
                        label="Choose Sparse Encoder",
                        value="MLM encoder (SPLADE-cocondenser-distil)"
                    )
                    input_text = gr.Textbox(
                        lines=5,
                        label="Enter your query or document text here:",
                        placeholder="e.g., Why is Padua the nicest city in Italy?"
                    )

                    # Custom Share Button and URL display
                    with gr.Row(elem_classes="share-button-container"):
                        share_button = gr.Button(
                            "🔗 Get Share Link",
                            elem_classes="custom-share-button",
                            size="sm"
                        )

                    share_output = gr.Textbox(
                        label="Share URL",
                        interactive=True, # Make it interactive so user can copy
                        visible=False,   # Start as hidden
                        placeholder="Click 'Get Share Link' to generate URL..."
                    )

                    info_output_display = gr.Markdown(
                        value="",
                        label="Vector Information",
                        elem_id="info_output_display"
                    )
                with gr.Column(scale=2): # Right column for the main representation output
                    main_representation_output = gr.Markdown()

            # Connect share button.
            # When share_button is clicked, get_current_share_url is called to populate share_output
            # and then share_output is made visible.
            share_button.click(
                fn=get_current_share_url,
                outputs=share_output
            ).then(
                fn=lambda: gr.update(visible=True),
                outputs=share_output
            )


            # Connect the core prediction logic
            model_radio.change(
                fn=predict_representation_explorer,
                inputs=[model_radio, input_text],
                outputs=[main_representation_output, info_output_display]
            )
            input_text.change(
                fn=predict_representation_explorer,
                inputs=[model_radio, input_text],
                outputs=[main_representation_output, info_output_display]
            )

            # Initial call to populate on load (optional, but good for demo)
            demo.load(
                fn=lambda: predict_representation_explorer(model_radio.value, input_text.value),
                outputs=[main_representation_output, info_output_display]
            )


        with gr.TabItem("Compare Encoders"):
            gr.Markdown("### Calculate Dot Product Similarity Between Encoded Query and Document")

            model_choices = [
                "MLM encoder (SPLADE-cocondenser-distil)",
                "MLP encoder (SPLADE-v3-lexical)",
                "Binary Bag-of-Words"
            ]

            gr.Interface(
                fn=calculate_dot_product_and_representations_independent,
                inputs=[
                    gr.Radio(
                        model_choices,
                        label="Choose Query Encoding Model",
                        value="MLM encoder (SPLADE-cocondenser-distil)"
                    ),
                    gr.Radio(
                        model_choices,
                        label="Choose Document Encoding Model",
                        value="MLM encoder (SPLADE-cocondenser-distil)"
                    ),
                    gr.Textbox(
                        lines=3,
                        label="Enter Query Text:",
                        placeholder="e.g., famous dishes of Padua"
                    ),
                    gr.Textbox(
                        lines=5,
                        label="Enter Document Text:",
                        placeholder="e.g., Padua's cuisine is as famous as its legendary University."
                    )
                ],
                outputs=gr.Markdown(),
                allow_flagging="never"
            )

# This block ensures the share URL is captured when the app launches
if __name__ == "__main__":
    # Launch and capture the share URL
    # Setting live=False for better control over when functions run
    # and to ensure the share_url is available early.
    launched_demo = demo.launch(share=True)
    # The share_url is typically available on the launched_demo object
    # This might vary slightly based on Gradio version.
    # For newer Gradio versions (>=3.x), the share_url is usually printed to console
    # and can sometimes be accessed via launched_demo.share_url directly if `return_url=True`
    # (though that was removed in recent versions as it's default now)

    # In older Gradio, you might need to manually copy the URL from console.
    # In newer Gradio, the share URL is often automatically put into a share button/link
    # in the UI (usually top right or footer) when share=True.
    # For our custom button, we're assuming we'll get the URL once it's launched.
    # Gradio automatically injects the share URL into the client-side JavaScript.
    # A simple way to get it from within the app is using Javascript or by storing it globally.

    # To bridge the gap, we'll tell the user to copy from the console for now,
    # or the built-in Gradio share icon (usually on the top right) will work automatically.

    print("\n--- Gradio App Launched ---")
    print("If a public share link is generated, it will be displayed in your console.")
    print("You can also use the '🔗 Get Share Link' button on the 'Sparse Representation' tab.")
    print("---------------------------\n")

    # In a real deployed scenario, this global_share_url might be populated by
    # a Gradio callback for `launch`, but for typical local execution,
    # the URL is primarily given via console/built-in UI.
    # The `get_current_share_url` function will return 'Share URL not available yet.'
    # until a share URL is actually generated by Gradio's backend when running `share=True`.
    # To truly populate `global_share_url` reliably for the custom button,
    # you often need to manually copy it or rely on a more advanced Gradio feature (like a custom JS callback or an API endpoint)
    # which is beyond simple `gr.Blocks` component connections.
    # The most direct approach for users is the default Gradio share icon/text.