app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForTokenClassification
from transformers import pipeline
import json


tokenizer = AutoTokenizer.from_pretrained("Babelscape/wikineural-multilingual-ner")
model = AutoModelForTokenClassification.from_pretrained("Babelscape/wikineural-multilingual-ner")

nlp = pipeline("ner", model=model, tokenizer=tokenizer, aggregation_strategy="simple")

def group_cat(entities):
    categories = {}
    for item in entities:
        group = item.get('entity_group')
        if group not in categories:
            categories[group] = [item]
        else:
            categories[group].append(item)
    return categories




def ner(text: str) -> str:
    """
    Searches the input text for named entities and returns them organized by category.
    
    Args:
        text (str): The input text to analyze.
    
    Returns:
        str: A json string representing dictionary where each key is a named entity category (e.g., 'PER', 'ORG', 'LOC', etc.), and the corresponding value is a list of entities found in the text under that category.
    """
    max_len = tokenizer.model_max_length
    stride = 50

    # Tokenizza con overflow per gestire testi lunghi
    inputs = tokenizer(
        text,
        return_overflowing_tokens=True,
        stride=stride,
        max_length=max_len,
        truncation=True,
        return_offsets_mapping=True,
        padding=False
    )

    all_entities = []
    seen = set()  # Per deduplicare (word, start, end)

    for input_ids in inputs["input_ids"]:
        chunk_text = tokenizer.decode(input_ids, skip_special_tokens=True)
        chunk_entities = nlp(chunk_text)

        for ent in chunk_entities:
            key = (ent["word"], ent["start"], ent["end"])
            if key not in seen:
                seen.add(key)
                all_entities.append(ent)

    
    ner_results =group_cat(all_entities)

    cleaned = {}
    for category, items in ner_results.items():
        cleaned[category] = {}
        for ent in items:
            cleaned[category][ent["word"]] = float(ent["score"])
    
    dict_ner = json.dumps(cleaned, indent=2, separators=(',', ': '), ensure_ascii=False)
    return dict_ner

# Create a standard Gradio interface
demo = gr.Interface(
    fn=ner,
    inputs=["text"],
    outputs="text",
    title="NER",
    description="Detect named entity within the text in input using the model Babelscape/wikineural - This interface works as MCP server as well."
)

# Launch both the Gradio web interface and the MCP server
if __name__ == "__main__":
    demo.launch(mcp_server=True)