app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForMaskedLM, AutoModel
import torch

# --- Model Loading ---
tokenizer_splade = None
model_splade = None
tokenizer_splade_lexical = None
model_splade_lexical = None
tokenizer_splade_doc = None # New tokenizer for SPLADE-v3-Doc
model_splade_doc = None     # New model for SPLADE-v3-Doc

# 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() # Set to evaluation mode for inference
    print("SPLADE v3 (cocondenser) model loaded successfully!")
except Exception as e:
    print(f"Error loading SPLADE (cocondenser) 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() # Set to evaluation mode for inference
    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 (NEW)
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)
    model_splade_doc.eval() # Set to evaluation mode for inference
    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 (still needed for splade-v3-lexical) ---
def create_lexical_bow_mask(input_ids, vocab_size, tokenizer):
    """
    Creates a binary bag-of-words mask from input_ids,
    zeroing out special tokens and padding.
    """
    bow_mask = torch.zeros(vocab_size, device=input_ids.device)
    meaningful_token_ids = []
    for token_id in input_ids.squeeze().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:
        bow_mask[list(set(meaningful_token_ids))] = 1
            
    return bow_mask.unsqueeze(0)


# --- Core Representation Functions ---

def get_splade_representation(text):
    if tokenizer_splade is None or model_splade is None:
        return "SPLADE (cocondenser) 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()
    else:
        return "Model output structure not as expected for SPLADE (cocondenser). 'logits' not found."

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

    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 = "SPLADE (cocondenser) Representation (All Non-Zero Terms):\n"
    if not sorted_representation:
        formatted_output += "No significant terms found for this input.\n"
    else:
        for term, weight in sorted_representation:
            formatted_output += f"- **{term}**: {weight:.4f}\n"

    formatted_output += "\n--- Raw SPLADE Vector Info ---\n"
    formatted_output += f"Total non-zero terms in vector: {len(indices)}\n"
    formatted_output += f"Sparsity: {1 - (len(indices) / tokenizer_splade.vocab_size):.2%}\n"

    return formatted_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()
    else:
        return "Model output structure not as expected for SPLADE v3 Lexical. 'logits' not found."

    # --- Apply Lexical Mask (always applied for this function now) ---
    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
    # --- End Lexical Mask Logic ---

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

    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 = "SPLADE v3 Lexical Representation (All Non-Zero Terms):\n"
    if not sorted_representation:
        formatted_output += "No significant terms found for this input.\n"
    else:
        for term, weight in sorted_representation:
            formatted_output += f"- **{term}**: {weight:.4f}\n"

    formatted_output += "\n--- Raw SPLADE Vector Info ---\n"
    formatted_output += f"Total non-zero terms in vector: {len(indices)}\n"
    formatted_output += f"Sparsity: {1 - (len(indices) / tokenizer_splade_lexical.vocab_size):.2%}\n"

    return formatted_output


# NEW: Function for SPLADE-v3-Doc representation (Binary Sparse)
def get_splade_doc_representation(text):
    if tokenizer_splade_doc is None or model_splade_doc is None:
        return "SPLADE v3 Doc model 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(model_splade_doc.device) for k, v in inputs.items()}

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

    if not hasattr(output, "logits"):
        return "SPLADE v3 Doc model output structure not as expected. 'logits' not found."

    # For SPLADE-v3-Doc, the output is often a binary sparse vector.
    # We will assume a simple binarization based on a threshold or selecting active tokens.
    # A common way to get "binary" is to use softplus and then binarize, or directly binarize max logits.
    # Given the "no weighting, no expansion" request, we'll aim for a strict presence check.

    # Option 1: Binarize based on softplus output and threshold (similar to UNICOIL)
    # This might still activate some "expanded" terms if the model predicts them strongly.
    # transformed_scores = torch.log(1 + torch.exp(output.logits)) # Softplus
    # splade_vector_raw = torch.max(transformed_scores * inputs['attention_mask'].unsqueeze(-1), dim=1).values
    # binary_splade_vector = (splade_vector_raw > 0.5).float() # Binarize

    # Option 2: Rely on the original BoW for terms, with 1 for presence
    # This aligns best with "no weighting, no expansion"
    vocab_size = tokenizer_splade_doc.vocab_size
    binary_splade_vector = create_lexical_bow_mask(
        inputs['input_ids'], vocab_size, tokenizer_splade_doc
    ).squeeze()
    
    # We set values to 1 as it's a binary representation, not weighted
    indices = torch.nonzero(binary_splade_vector).squeeze().cpu().tolist()
    if not isinstance(indices, list): # Handle case where only one non-zero index
        indices = [indices] if indices else [] # Ensure it's a list even if empty or single
    
    # Values are all 1 for binary representation
    values = [1.0] * len(indices) 
    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 binary

    formatted_output = "SPLADE v3 Doc Representation (Binary Sparse - Lexical Only):\n"
    if not sorted_representation:
        formatted_output += "No significant terms found for this input.\n"
    else:
        # Display as terms with no weights as they are binary (value 1)
        for i, (term, _) in enumerate(sorted_representation):
            # Limit display for very long lists for readability
            if i >= 50:
                formatted_output += f"...and {len(sorted_representation) - 50} more terms.\n"
                break
            formatted_output += f"- **{term}**\n"

    formatted_output += "\n--- Raw Binary Sparse Vector Info ---\n"
    formatted_output += f"Total activated terms: {len(indices)}\n"
    formatted_output += f"Sparsity: {1 - (len(indices) / tokenizer_splade_doc.vocab_size):.2%}\n"

    return formatted_output


# --- Unified Prediction Function for Gradio ---
def predict_representation(model_choice, text):
    if model_choice == "SPLADE (cocondenser)":
        return get_splade_representation(text)
    elif model_choice == "SPLADE-v3-Lexical":
        # Always applies lexical mask for this option
        return get_splade_lexical_representation(text)
    elif model_choice == "SPLADE-v3-Doc": # Simplified to a single option
        # This function now intrinsically handles binary, lexical-only output
        return get_splade_doc_representation(text) 
    else:
        return "Please select a model."

# --- Gradio Interface Setup ---
demo = gr.Interface(
    fn=predict_representation,
    inputs=[
        gr.Radio(
            [
                "SPLADE (cocondenser)",
                "SPLADE-v3-Lexical",
                "SPLADE-v3-Doc" # Only one option for Doc model
            ],
            label="Choose Representation Model",
            value="SPLADE (cocondenser)" # Default selection
        ),
        gr.Textbox(
            lines=5,
            label="Enter your query or document text here:",
            placeholder="e.g., Why is Padua the nicest city in Italy?"
        )
    ],
    outputs=gr.Markdown(),
    title="🌌 Sparse Representation Generator",
    description="Enter any text to see its sparse vector representation.", # Simplified description
    allow_flagging="never"
)

# Launch the Gradio app
demo.launch()