app.py

#########################################################################
# Copyright (C)                                                       	#
# 2025-August Sen Li (Sen.Li.Sprout@gmail.com)		     			    #
# Permission given to modify the code only for Non-Profit Research		#
# as long as you keep this declaration at the top 						#
#########################################################################
import os

import gradio as gr
import huggingface_hub
import sentence_transformers

from transformers import AutoTokenizer, AutoModel
import torch

# ----------------------------------------------------------------------------------------------------------------------
def func_ClearInputs():        return "", "", ""


# str_ModelID_ClinicalBERT = "medicalai/ClinicalBERT"
str_ModelID_ModernBERT = "answerdotai/ModernBERT-large"

# str_ModelID_ClinicalBERT = "TsinghuaC3I/Llama-3-8B-UltraMedical"

# tokenizer = AutoTokenizer.from_pretrained(str_ModelID_ModernBERT)
# model_ClinicalBERT = AutoModel.from_pretrained(str_ModelID_ModernBERT)

# Wrap ClinicalBERT inside SentenceTransformers
word_embedding_model = sentence_transformers.models.Transformer(str_ModelID_ModernBERT)
pooling_model = sentence_transformers.models.Pooling(word_embedding_model.get_word_embedding_dimension())
sentenceModel_ModernBERT = sentence_transformers.SentenceTransformer(modules=[word_embedding_model, pooling_model])
                                                               # device="cuda")


def get_SentenceEmbeddings_ModernBERT(sentence):
    # Encode sentences in batches efficiently
    embeddings = sentenceModel_ModernBERT.encode(sentence)

    return embeddings


def get_sentence_embedding(sentence: str) -> torch.Tensor:
    # Tokenize and encode
    inputs = tokenizer(sentence, return_tensors="pt", truncation=True, padding=True)

    # Get hidden states
    with torch.no_grad():
        outputs = sentenceModel_ModernBERT(**inputs)

    # outputs.last_hidden_state shape: (batch_size, seq_len, hidden_dim)
    token_embeddings = outputs.last_hidden_state

    # Create sentence embedding (mean pooling across tokens, ignoring padding)
    attention_mask = inputs['attention_mask']
    mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    sentence_embedding = torch.sum(token_embeddings * mask_expanded, dim=1) / torch.clamp(mask_expanded.sum(dim=1),
                                                                                          min=1e-9)

    return sentence_embedding.squeeze()

def func_sBERT_SimilarityResult(str_Text_1, str_Text_2):
    if not str_Text_1.strip() or not str_Text_2.strip():
        return "Both text inputs must be non-empty."

    # 01. Load SBERT model (you can choose other pre-trained models too)
    inferenceClient = huggingface_hub.InferenceClient(provider="hf-inference")

    # 02. Get sentence embeddings
    # str_ModelID_sBERT = "medicalai/ClinicalBERT"
    # arrEmbedding_Text_1 = inferenceClient.feature_extraction(text=str_Text_1, model=str_ModelID_sBERT)
    # arrEmbedding_Text_2 = inferenceClient.feature_extraction(text=str_Text_2, model=str_ModelID_sBERT)

    # arrEmbedding_Text_1 = get_sentence_embedding(str_Text_1)
    # arrEmbedding_Text_2 = get_sentence_embedding(str_Text_2)

    arrEmbedding_Text_1 = get_SentenceEmbeddings_ModernBERT(str_Text_1)
    arrEmbedding_Text_2 = get_SentenceEmbeddings_ModernBERT(str_Text_2)

    # 03. Compute cosine similarity
    tensor_Similarity = sentence_transformers.util.pytorch_cos_sim(arrEmbedding_Text_1, arrEmbedding_Text_2)
    f_Similarity = tensor_Similarity.item()

    return f"Clinical Similarity Score: {f_Similarity:.4f}"

# ----------------------------------------------------------------------------------------------------------------------
# Launch the interface and MCP server
if __name__ == "__main__":
    print(f"os.getcwd() = {os.getcwd()}")
    os.system(f"echo ls -al {os.getcwd()} && ls -al {os.getcwd()}")
    os.system(f"echo ls -al /: && ls -al /")
    os.system(f"echo ls -al /home/: && ls -al /home/")

    # 03. Gradio UI elements
    with gr.Blocks() as grBlocks_SentenceSimilarity__MCP_Server:
        gr.Markdown("# ModernBERT for Clinical Text Similarity using HF Inference Server, MaxSeqLength==8192")
        gr.Markdown("This application calculates Cosine Similarity Score between two Texts' ModernBERT Sentence-Embeddings")

        with gr.Row():
            grTextBox_Input_1 = gr.Textbox(label="Text Panel 1", lines=20)
            grTextBox_Input_2 = gr.Textbox(label="Text Panel 2", lines=20)

        with gr.Row():
            with gr.Column(scale=1):
                grButton_Clear  = gr.Button("Clear")
                grButton_Submit = gr.Button("Submit")
            with gr.Column(scale=3):
                grTextbox_Output = gr.Textbox(label="Similarity Result", interactive=False)

        # Set button functionality
        grButton_Submit.click(fn=func_sBERT_SimilarityResult, inputs=[grTextBox_Input_1, grTextBox_Input_2], outputs=grTextbox_Output)
        grButton_Clear.click(fn=func_ClearInputs, inputs=[], outputs=[grTextBox_Input_1, grTextBox_Input_2, grTextbox_Output])


    # 04. Launch Gradio MCP server
    grBlocks_SentenceSimilarity__MCP_Server.launch(mcp_server=True, share=True)