app.py

import torch
import gradio as gr
from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification
from rapidfuzz import process

# ------------------ Load models once ------------------
# Sentiment model
sentiment_pipeline = pipeline(
    "sentiment-analysis",
    model="sreejith8100/indian_output",
    tokenizer="sreejith8100/indian_output",
    device=0 if torch.cuda.is_available() else -1
)

# NER model
ner_tokenizer = AutoTokenizer.from_pretrained("ai4bharat/IndicNER", use_fast=True)
ner_model = AutoModelForTokenClassification.from_pretrained("ai4bharat/IndicNER")
ner_pipeline = pipeline(
    "ner",
    model=ner_model,
    tokenizer=ner_tokenizer,
    aggregation_strategy="simple",
    device=0 if torch.cuda.is_available() else -1
)

# Canonical entity list
CANONICAL_ENTITIES = [
    "V Abdurahiman / വി അബ്ദുറഹിമാൻ",
    "P A Mohamed Riyas / പി എ മുഹമ്മദ് റിയാസ്",
    "P Rajeev / പി രാജീവ്",
    "Saji Cherian / സജി ചെറിയാൻ",
    "Roshy Augustine / റോഷി ഓഗസ്റ്റിൻ",
    "R Bindu / ആർ ബിന്ദു",
    "A K Saseendran / എ കെ സസീന്ദ്രൻ",
    "O R Kelu / ഒ ആർ കെലു",
    "J Chinchurani / ജെ ചിഞ്ചുറാണി",
    "K N Balagopal / കെ എൻ ബാലഗോപാൽ",
    "K Krishnankutty / കെ കൃഷ്ണൻകുട്ടി",
    "Veena George / വീണാ ജോർജ്",
    "Antony Raju / ആന്റണി രാജു",
    "K Rajan / കെ രാജൻ",
    "M B Rajesh / എം ബി രാജേഷ്",
    "Chittayam Gopakumar / ചിറ്റയം ഗോപകുമാർ",
    "K Radhakrishnan / കെ രാധാകൃഷ്ണൻ",
    "Pinarayi Vijayan / പിണറായി വിജയൻ",
    "V Sivankutty / വി ശിവൻകുട്ടി",
    "K K Shailaja / കെ കെ ശൈലജ"
]

def map_entity(entity_text, known_entities=CANONICAL_ENTITIES, threshold=70):
    match, score, _ = process.extractOne(entity_text, known_entities)
    if score >= threshold:
        return match
    return None

# Map raw model labels to readable ones
label_map = {
    "LABEL_0": "POSITIVE",
    "LABEL_1": "NEGATIVE",
    "LABEL_2": "NEUTRAL"
}

# ------------------ Prediction function ------------------
def predict(sentence):
    # Run sentiment
    sent_pred = sentiment_pipeline(sentence)[0]
    human_label = label_map.get(sent_pred["label"], sent_pred["label"])  # map it

    # Run NER + map
    entities = ner_pipeline(sentence)
    mapped_entities = [map_entity(ent["word"]) for ent in entities if map_entity(ent["word"])]

    return {
        "sentence": sentence,
        "prediction": human_label,   # use mapped label
        "score": float(sent_pred["score"]),
        "mapped_entities": list(set(mapped_entities))
    }

# ------------------ Gradio Interface ------------------
demo = gr.Interface(
    fn=predict,
    inputs=gr.Textbox(label="Enter a sentence"),
    outputs=gr.JSON(label="Result"),
    title="Entity + Sentiment Analysis",
    description="Upload a sentence in Malayalam/English. The app detects entities and predicts sentiment."
)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860)