handler.py

from typing import Dict, List
from transformers import AutoTokenizer
from adapters import AutoAdapterModel

class AdapterHandler:
    def __init__(self):
        self.tokenizer = None
        self.model = None
        
    def initialize(self, model_dir: str):
        self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
        self.model = AutoAdapterModel.from_pretrained(model_dir)
        # Load adapter - adjust path as needed
        self.model.load_adapter("specter2_proximity", source="local")
        self.model.set_active_adapters("specter2_proximity")
        return self

    def __call__(self, data: Dict[str, List[str]]) -> Dict[str, List[float]]:
        titles = data.get("title", [""])
        abstracts = data.get("abstract", [""] * len(titles))
        
        # Combine inputs
        combined = [
            title + self.tokenizer.sep_token + (abstract or "")
            for title, abstract in zip(titles, abstracts)
        ]
        
        # Tokenize
        inputs = self.tokenizer(
            combined,
            padding=True,
            truncation=True,
            return_tensors="pt",
            return_token_type_ids=False,
            max_length=512
        )
        
        # Get embeddings
        with torch.no_grad():
            outputs = self.model(**inputs)
            embeddings = outputs.last_hidden_state[:, 0, :].numpy()
        
        return {"embeddings": embeddings.tolist()}